FutureG Research Meetings

November 18

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 
Subtitle 1: M-AFTER: Learning Specialized Teams for Multi-Agent Reasoning
Presenter: Hongbo Li, postdoc researcher, Ohio State
 
Subtitle 2: A Super-action Lyapunov Drift Approach for Optimizing Age of Information
Presenter: Tasmeen Zaman Ornee, postdoc researcher, Ohio State
 

Abstract: Effective collaboration requires timely information and specialized capabilities. This talk examines information scheduling and agent learning in collaborative systems. For semi-decentralized systems with probabilistic information sharing, we formulate weighted sum Age of Information (AoI) minimization as a restless multi-armed bandit problem under resource and scheduling constraints. We establish indexability and asymptotic optimality within a super-action framework, then incorporate Lyapunov drift and a transmission penalty to develop a low-complexity Maximum Lyapunov-based Super-action index (MLS) policy with closed-form indices and performance guarantees. To support collaboration across diverse reasoning tasks, we introduce M-AFTER, which combines specialization by workflow role and problem type through reusable expert bundles. Staged training develops these bundles using outcome rewards and role-specific feedback, then trains a router over the frozen experts. Simulations show that MLS reduces normalized AoI relative to Maximum Age First and randomized policies. Experiments on Reasoning Gym and mixed code, mathematics, and science tasks show that M-AFTER outperforms the evaluated single-agent and multi-agent baselines.

November 4

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola

Topic: Agent-Coordinated ISAC Using Multi-objective Optimization and Reinforcement Learning
Presenter: Abigail Eastman, PhD student

Abstract: Future-generation integrated sensing and communications (ISAC) systems
must balance competing objectives while sharing spectrum, allocating
limited resources, and adapting to time-varying environments.
Building on our group’s prior work using Q-learning to improve radar
performance in time-varying environments, we propose a decision-theoretic
framework for agent-coordinated ISAC. The framework combines multi-objective
waveform co-design with reinforcement learning for adaptive resource allocation
for radar tracking and multi-user communications. Multi-objective optimization
identifies Pareto-optimal waveform pairs that trade off between radar
target state estimation error and communications bit error rate.
A Q-learning agent selects the transmit waveform pair and next system power
resource modes based on current signal-to-noise ratios, mission priority,
and available resources. The agent learns a coordination policy that minimizes
the expected cumulative cost which combines radar and communications performance
metrics with a penalty cost for failing to meet mission-priority requirements.

October 28, 2026

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: Opportunities and Challenges of Consumer-Grade Unmanned Aerial Vehicles in FutureG Networks
Presenter: Andrew Herschfelt, Research Professor, ASU

Abstract: Unmanned aerial vehicles (UAVs) have quickly become accessible and abundant in urban civilian environments. These platforms enable a wide range of novel and powerful applications for FutureG networks, but also present emergent technical and logistical challenges. We explore the opportunities that UAVs enable in FutureG networks and the implications of introducing numerous wireless devices into already cluttered environments. To characterize the impact these devices have on the physical and electromagnetic environments, we explore a range of characterization techniques, including radar cross-section (RCS), acoustic and electromagnetic micro-Doppler signatures, and passive detection and localization using telemetry signals. We leverage the Electromagnetic Anechoic Chamber (EMAC) at Arizona State University to conduct high-precision measurements on these devices in the range of 0 - 20 GHz, and explore alternate approaches for conducting similar measurements in the higher frequency ranges supported by emerging FutureG hardware.

October 7, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: CHORUS: Persistent Backdoor Attacks in Federated Multimodal Sensing
Presenter: Ruobing Li, PhD student, Arizona State

Abstract: Federated multimodal sensing (FedMMS) trains models over complementary sensing modalities without centralizing privacy-sensitive data. It inherits federated learning's vulnerability to backdoor attacks, where the difficulty lies less in injecting a backdoor than in sustaining it against repeated dilution by honest clients. We show that multimodal fusion supplies exactly what such an attacker needs: because modality-specific encoders meet only at fusion, a target defined in the fused representation is tied to no single modality and is computed identically by every compromised client, whatever sensors it carries. Optimizing toward it drives compromised updates into agreement while heterogeneous benign updates cancel under averaging.

We present CHORUS, the first backdoor attack against FedMMS, which reconstructs a round-adaptive target prototype from the released global model and optimizes modality-specific perturbations under a blockwise gradient-alignment objective, never manipulating uploaded updates. Under measurable conditions, coordinated contributions accumulate linearly while benign fluctuation grows only as O(√T), with scarce modalities compromised soonest. Experiments on federated multimodal human activity recognition show that CHORUS sustains 94.6% attack success over 100 rounds with 10% compromised clients, preserves clean accuracy, survives compromised-client withdrawal, and withstands two robust aggregation defenses.


 

September 30, 2026

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Digital Twin Model for RIS-aided Device-Free Sensing of Indoor Static Object Localization
Presenter: Qifeng Chen, PhD student

Abstract: We consider the problem of indoor static target detection/localization for the scenario where the transmitter (TX) and target are inside the room and the receiver (RX) is outside. Earlier we had shown how reconfigurable intelligent surface (RIS) beam scanning can be used to create controllable reflection paths and use it to infer the object direction. Our analytical model was calibrated using real-world experimental measurements. Since real-world data collection is extremely time-consuming, in this presentation, we describe the digital twin (DT) model that we developed for this set up. We utilized the NVIDIA Sionna RT framework to build the DT model by integrating a geometry-based 3D indoor model, transmitter and receiver antenna models, a low-poly 3D human model, and a customized near-field RIS model. We generated data corresponding to different human object locations and trained light-weight MLP and CNN models using this data to predict the angle subtended by the human object to the RIS. We find that both models achieve only around 60.0% accuracy when evaluated on real measurement data. However, after fine-tuning with real data, the accuracy increased to 93.3% for MLP and 88.7% for CNN. Our next step is to extend this system to accurately predict the location of the human object using triangulation in a two-RIS system.

Topic 2: Compilation framework for parallel programming of multi-tiled CGRA architectures
Presenter: Anish Vipperla, PhD student

Abstract: Coarse-Grained Reconfigurable Array (CGRA) architectures offer high-performance acceleration of compute-intensive spatially mapped workloads while ensuring high flexiblity in terms of reconfigurability and domain adaptation. They offer an excellent trade-off between performance of a fixed-function ASIC and fine-grained reconfigurability of an FPGA. The performance of a CGRA-based accelerator heavily depends on the software-stack that supports compilation and execution of compute-intensive kernels on the reconfigurable fabric. Existing works fall short of utilizing the resources effectively for large homogeneous PE arrays or tiled array-architectures. Further, they lack support for standardized programming frontends and execution models, requiring architecture-specific directives in the high-level code. In this presentation, we showcase our work on the development of a compilation framework that supports standardized parallel programming frontend and parallel execution paradigm for improved resource-utilization and higher-throughput on a multi-processor and multi-chiplet tiled-CGRA architecture. Our framework offers benefits of parallel programming and execution model through thread-level, tile-level, process-level, kernel, and chiplet-level parallelism. Our hardware-software tiled-CGRA stack is particularly suited for adaptive signal processing and wireless communication applications that require fast and dynamic workload switching in changing environments by offering high-throughput and fast-reconfiguration.


 

September 23, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: VibPrint: Acoustic Fingerprinting via mmWave-Sensed Acoustic Vibrations
Presenter: Sisi Chen, PhD student, Arizona State

Abstract: Acoustic fingerprinting identifies mobile devices through the unique characteristics of their integrated speaker and microphone. Existing approaches extract fingerprints from microphone-recorded acoustic signals. In this paper, we introduce Acoustic Vibration Fingerprints (AVFs), a complementary fingerprinting primitive that characterizes devices through their mechanical vibration responses to acoustic excitation. We present VibPrint, the first system to authenticate mobile devices using mmWave-sensed AVFs. Leveraging a commodity FMCW mmWave radar, VibPrint captures speaker-induced surface vibrations, constructs three complementary AVFs based on single-tone sweeps, dual-tone pairs, and wideband chirps, and learns robust fingerprint representations for prototype-based authentication. We evaluate VibPrint on 14 commercial smartphones and 60 homogeneous microspeakers. Experimental results demonstrate up to 100% authentication accuracy while maintaining robust performance across different sensing distances, device orientations, and deployment environments. These results establish acoustic vibration fingerprinting as a practical and complementary modality for mobile device authentication.


 

September 16, 2026

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Resource-Aware ISAC Intrusion Detection in Tactical Networks: A Stackelberg Game Approach
Presenter: Xuanli Lin, Ph.D. student

Abstract: Modern wireless systems increasingly rely on spectrum sharing to improve spectrum utilization and support heterogeneous users and applications. In mission-oriented settings, however, different transmission requests may have strict priority relationships: a high priority request should not be displaced simply because many lower priority requests provide greater aggregate reward. Existing tiered access systems such as the Citizens Broadband Radio Service (CBRS) motivate this setting, but do not directly address how application-level transmission requests should be scheduled across limited wireless resources while respecting strict priorities.

To address this challenge, we study a tiered wireless resource allocation problem in which transmission requests have different priority tiers, processing requirements, deadlines, and rewards, and must be assigned to a set of orthogonal channels. We use lexicographic reward maximization to enforce strict priority among tiers without requiring artificially chosen weight scales. We show that the resulting scheduling problem is NP hard even for a single tier and a single channel. For a fixed number of channels, we develop an exact pseudo polynomial dynamic programming algorithm that jointly tracks the load across all channels and returns a lexicographically optimal schedule.

To improve scalability, we develop two approximation algorithms based on an exact single channel dynamic programming subroutine. The Channel First algorithm assigns requests one channel at a time, while the Tier First algorithm processes priority tiers sequentially while preserving previously assigned higher priority requests. Both algorithms run in O(CKD) time and provide approximation guarantees of at least 1-1/e under the lexicographic objective. Numerical experiments show that these algorithms remain close to the exact solution on instances where exact computation is feasible, while scaling to substantially larger numbers of requests and channels. Together, these results provide a practical framework for scheduling wireless requests under strictly tiered priorities.


 

September 9, 2026

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Self-Evolving LLM Agents and Age-Penalty Scheduling with Provable Guarantees in Real-time Wireless Networks
Subtitle 1: Self-Evolving LLM Agents for Fast O-RAN Control with Provable Learning Guarantees
Presenter: Hongbo Li, postdoc researcher, Ohio State
Subtitle 2: A Super-action Lyapunov Drift Approach for Optimizing the Age Penalty in Semi-decentralized Systems
Presenter: Tasmeen Zaman Ornee, postdoc researcher, Ohio State

Abstract: One fundamental challenge in Real-time wireless control is that the most important decisions require computation that does not fit inside the control deadline. We present two results unified by a two-timescale principle, in which a slow loop distills expensive computation into a lightweight object that a fast loop evaluates within its deadline, with online correction preserving provable guarantees. First, for O-RAN control, we introduce SPECTRA, whose Non-RT loop distills a semantic model into an inspectable table over contexts and feasible xApp active sets, while a Near-RT LpC-UCB controller corrects this prior from feedback and prices SLA violations without invoking an LLM, achieving sublinear regret under a linear residual model. Second, for semi-decentralized updating under general Age of Information penalties, we show that the correct Lyapunov function is the bias function of the reference chain. This yields a closed-form super-action index, computable in O(N2) per group, that is within a penalty-dependent factor of the optimal: two for linear age, six for quadratic.


 

September 2, 2026

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic 1: Spatial Time-Frequency Channel Modeling for Future Generation Wideband MIMO Systems
Presenter: Pinjia Zhang, PhD student

Abstract: In mmWave communications, the spatial time-frequency channel model maps signal propagation over highly directional scattering paths into a virtual statistical representation of the channel's spatial-temporal-spectral effect on the transmitted waveform. We extend the channel representation to characterize the frequency-dependent spatial variations inherent in future generation ultra-wideband communication systems. In such systems, the extremely large bandwidth induces frequency-dependent steering vectors, causing the beam squint effect. This effect deteriorates beamforming performance, as different signal frequencies are steered toward different spatial directions. We use the new channel representation to demonstrate the improvement in performance as it allows for the physical array response to change with frequency as well as the angle of arrival.

Topic 2: OTFS Modulation Simplified (short and as time permits)
Presenter: Antonia Papandreou-Suppappola, ASU

Abstract: The original authors of orthogonal time-frequency space (OTFS) modulation (Hadani, etc, 2017) explain its operating principles in terms of the channel's delay-Doppler impulse response and the twisted convolution relation between the channel input and output. We provide a different explanation, using time-frequency analysis principles and show that a similar modulation scheme was previously published by Molisch et al, 1998 and Sayeed et al, 2004.


 

August 12, 2026

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: Experimental Validation of RF Convergence Techniques
Presenter: Alex Chiriyath, Research Professor, ASU

Abstract: We aim to study how ISAC systems can operate with a true 5GNR waveform and base station. We employ the Amarisoft Callbox to deploy a real 5G base station in a laboratory environment and program Ettus B210 USRPs to act as a 5G receiver/relay and radar target emulator. The aim is to program a USRP to act as an ISAC receiver capable of disentangling the relayed 5G signal from the signal reflected by the target. We use this experimental testbed to validate simple RF convergence techniques in over-the-air experiments.


 

July 22, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: StegoBFI: A Covert Channel in 802.11 Beamforming Feedback
Presenter: Mingze Pan, PhD student

Abstract: Beamforming is a mandatory component of modern Wi-Fi (IEEE 802.11ac/ax/be). To support it, access points periodically initiate channel sounding, during which stations return compressed and unprotected Beamforming Feedback Information (BFI). We present StegoBFI, a steganographic covert channel that hides data in the BFI that 802.11 devices already exchange. A client encodes arbitrary data into the least significant bits of the BFI, leaving the feedback nearly indistinguishable from the original and the AP's beamforming performance unaffected — and thus undetectable. StegoBFI is inherently dual-use. Legitimate clients can piggyback data (e.g., continuous authentication messages) onto existing feedback without introducing extra packets that disrupt normal communication. Compromised clients can exploit the same channel to exfiltrate stolen data, bypassing on-device network traffic monitoring. In this talk, we describe the problem formulation, scheme design, and preliminary evaluation on commodity Wi-Fi devices.


 

July 15, 2026

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: RIS-aided Indoor Device-Free Sensing for Static Object Localization
Presenter: Qifeng Chen

Abstract: We consider the problem of indoor static target detection/localization for the scenario where the transmitter (TX) and target are inside the room and the receiver (RX) is outside. Unlike conventional device-free sensing methods that rely on temporal channel variations and therefore struggle with static targets, we propose to utilize reconfigurable intelligent surface (RIS) beam scanning to create controllable reflection paths and infer object direction from the received power differences. We develop a RIS scanning model that considers the effect of both main lobe and side lobes on the reflection paths. The model is calibrated using real-world experimental measurements collected with a 28 × 36-element RIS at 5.55 GHz and USRP X310 transceivers. For localization, we utilize a two-RIS configuration and estimate the object position through triangulation. We study the impact of room size and inter-RIS distance on localization coverage. Simulation and experimental results with real hardware show good agreement, with about 1 dB discrepancy at sampled points, and confirm the feasibility of RIS-aided static object detection and localization.

Topic 2: Enabling Real-time Rendering for XR
Presenter: Anish Vipperla

Abstract: Real-time rendering for Extended Reality (XR) experiences must deliver high frame rate and resolution while maintaining low frame noise and user-side latency. Real-time ray tracing in the XR device is particularly challenging because of its energy and resource-constraints. So we propose a cloud-based real-time rendering solution that offloads key components of the rendering pipeline to the cloud and performs post-processing tasks on the headset. We reduce the rendering time by rendering at low frame rate, resolution, and samples per pixel on the cloud and then addressing the quality degradation through post-processing tasks such as Frame Denoising, Super-Resolution, and Frame Interpolation on the headset. In this talk, we present our initial work on implementing denoising algorithms for resource-constrained devices. We find that while the state-of-the-art machine learning (ML) techniques offer superior output image quality, they often fail to meet the frame latency constraints imposed by real-time XR streaming. On the other hand, existing non-ML spatiotemporal filtering techniques have low frame latency but poor image quality. So we propose to add a lightweight ML-based enhancement module to the spatiotemporal filter. We demonstrate that such a technique results in visually acceptable output while operating within the strict per-frame time budget of real-time XR systems.


 

July 8, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: RobotAuth: Communication-Free Authentication of Robots via Physical Challenges and Responses
Presenter: Cong Wei, Ph.D. student, ASU

Abstract: Large open physical environments—smart cities, battlefields, disaster-response sites, industrial facilities—increasingly host many visually similar robots (humanoids, quadrupeds) operating side by side. In these settings, misidentifying a robot can trigger unsafe physical actions, leak mission-critical information, or disrupt coordinated operations. Adversaries can exploit visual similarity and intermittent connectivity to evade conventional network-based identity checks. We propose RobotAuth, a communication-free authentication framework that distinguishes legitimate robots from adversarial ones when digital identity channels are unavailable, compromised, or untrusted.

RobotAuth works by having the system operator embed micro-structural details—object-dependent physical challenges—into selected objects in the environment (e.g., road signs, trash bins, corridor markers). Any geometric region of an object, including surface textures, 3D contours, or structural edges, can serve as a carrier for encoded challenge information. When a legitimate robot perceives an instrumented object, it decodes the embedded challenge through its vision module and produces a physical response—a short, random sequence of robot actions—determined by its operator-defined authentication policy and unique secret key. A camera surveillance system connected to a backend server verifies this response and detects adversarial robots which cannot perform a correct physical-action sequence.

We formulate the embedding and decoding of physical challenges as a communication problem over a physical-visual channel, and propose an optimal encoding method that maximizes edge-encoding capacity subject to reliability, 3D-manufacturing, and unobtrusiveness constraints. We prototype RobotAuth using a Prusa MK4S 3D printer and a Unitree Go2 EDU quadruped robot, with promising initial results.


 

July 1, 2026

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Resource-Aware ISAC Intrusion Detection in Tactical Networks: A Stackelberg Game Approach
Presenter: Xuanli Lin, Ph.D. student, ASU

Abstract: Integrated sensing and communication (ISAC) enables wireless infrastructure to support both data communication and physical environment sensing, making it a promising approach for intrusion detection in tactical networks. In such environments, sensing and computational resources are limited, intruder behavior is uncertain, and different targets may have heterogeneous values and risks. These challenges make resource-aware sensing placement essential for protecting mission-critical assets against strategic attackers.

To address this challenge, we model ISAC-based intrusion detection as a Stackelberg security game on an undirected graph, where the defender first commits to a sensing and offloading strategy and the attacker then observes this strategy and selects one or more attack routes. For a fixed defender strategy, we show that the attacker’s best response can be computed efficiently by transforming route detection probabilities into additive edge weights and applying a shortest-path algorithm. However, the defender’s optimization problem remains computationally difficult due to the combinatorial sensing and offloading constraints.

To improve scalability, we develop resource-aware heuristic algorithms for computing effective defender strategies. Resource-Aware Downgrading starts from a fully loaded sensing configuration and progressively reduces sensing levels while minimizing the loss in defender utility. Local Exchange Refinement then improves a feasible strategy by exploring neighboring sensing configurations through local upgrades and exchanges. Together, these methods provide a practical framework for allocating limited ISAC resources against intelligent intruders while respecting budget and server capacity constraints.


 

June 24, 2026

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: LLM-Adapted Network Management and Scheduling for Minimizing Age-penalty in Mission-Critical Networks
Subtitle 1: Self-Evolving LLM Agents for Adaptive Network Management
Presenter: Hongbo Li, postdoc researcher, Ohio State
Subtitle 2: Minimizing Age-penalty over Wireless Networks: A Super-Action Lyapunov Drift Approach
Presenter: Tasmeen Z. Ornee, postdoc researcher, Ohio State

Abstract: In time-sensitive environments, such as autonomous UAV swarms, networked-controlled systems, and vehicle-to-everything (V2X) systems, successful communication relies on intelligent decision-making and low-latency data distribution. In this research, we present two frameworks to achieve semantic AI reasoning and performance optimization for adaptive network management. The first work proposes an LLM-prior constrained contextual bandit (LP-CUCB) framework that illustrates high-level semantic reasoning into rapid, auditable control actions. Through this approach, this framework enables the system to adaptively manage complex network tasks, including routing and resource allocation. In the second work, we complement this management layer with a Super-action Framework that enables partial decomposition of complex scheduling tasks (RMAB with additional hard constraint) into group-level sub-problems. By utilizing a Lyapunov drift plus penalty-based scheduling policy, we minimize the long-term age penalty across these groups under hard channel resource constraints. Our results demonstrate that both of these frameworks combine semantic agent reasoning with scheduling guarantees and provide a robust, low-latency solution for maintaining better performance in dynamic and mission-critical scenarios.


 

June 10, 2026

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Learning to Track: A Decision-Theoretic Framework for Time-Varying Environments
Presenter: Yongho Seo, PhD Candidate (defense June 23)

Abstract: The presentation highlights the candidate's research contributions supported under the DoD Center of Excellence in Future Generation Wireless Technology. The research formulated novel computational methods under a decision-theoretic framework to overcome core challenges of next-generation systems in tracking multiple targets in dynamic complex environments. An example of such a highly complex scenario involves a counter-measure operation, where a 6G base station must track the trajectories of incoming micro-drone swarms to simultaneously neutralize multiple targets. The research contributed methods to (a) sequentially learn target identities, measurement associations and time-varying noise characteristics in dynamic environments with clutter; (b) adapt tracking parameters to manage multiple tracks under time-varying conditions; (c) learn and resolve track ambiguities due to limited sensor resolution and high interference from co-existing systems sharing spectrum; and (d) model sensor systematic time-varying error functions in multimodal sensing systems to improve measurement-to-target association.


 

June 3, 2026

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: ISaC: Challenges and Technologies
Presenter: Prof. Dan Bliss, ASU

Abstract: Historically, sensing and communications have been developed and implemented, separately. Given the technical difficulty in getting any system to work reliably, this separation, while often theoretically suboptimal, was almost certainly the correct approach. However, because of the growing capabilities in RF frontend technologies and computational systems, the prospect of multiple-function RF approaches is becoming viable. I provide a conceptual and theoretical framework for the more general problem of multiple-function RF systems (RF convergence) and then focus on integrated sensing and communications (ISaC) systems that are typically associated with implementations of radar and communications using the next generation of cellular systems. I provide an overview of the set of challenges technical and economic that must be addressed to enable broad adoption and provide the greatest benefit from ISaC approaches. I broadly discuss motivations, challenges, and technologies for ISaC. Topics include radar performance, waveform optimization, algorithmic options, and advanced computational approaches.


 

May 13, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: No Where to Hide: Reconfigurable Intelligent Surface-Assisted Hidden Device Detection and Localization
Presenter: Yan Zhang, Ph.D. student

Abstract: Detecting hidden WiFi surveillance devices in large indoor environments is increasingly difficult as attackers adopt MAC address randomization and motion-triggered streaming behaviors. Prior detection approaches rely on stable MAC addresses, which can be readily obfuscated by modern attackers, breaking the association between packets and physical transmitters. We present RISLoc, the first fully autonomous system for detecting and localizing hidden WiFi-based surveillance devices under such adversarial conditions. RISLoc mounts a reconfigurable intelligent surface (RIS) on a mobile robot, using the robot's motion as an active stimulus to trigger motion-responsive devices, while the RIS performs directional RIS beamscanning for spatial localization. RISLoc addresses four key challenges: MAC obfuscation, reliable device detection, RIS coverage identification, and robust device localization. Specifically, RISLoc progressively identifies and localizes hidden devices through clustering-based MAC de-randomization, motion-triggered traffic pattern matching, RIS-induced coverage filtering, multi-position RIS beamscanning, and optimization-based triangulation. Extensive prototype-based experiments demonstrate that RISLoc effectively detects hidden devices and achieves accurate localization under diverse conditions.


 

April 29, 2026

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Federated Learning on Satellite Constellations
Presenter: Ziyi Zhao, Ph.D. student

Abstract: Low Earth Orbit (LEO) satellite constellations provide a scalable platform for learning models pertaining to applications such as weather monitoring and disaster prediction, with satellites collecting geographically distributed data. Directly applying conventional federated learning (FL) to LEO systems faces significant challenges due to intermittent satellite-ground connectivity, short communication windows, non-IID data, and limited onboard energy. To address these challenges, we propose a novel FL framework tailored for LEO satellite constellations. At the intra-orbit level, satellites leverage inter-satellite links to aggregate local models and form an orbit-level model. The orbit-level models are sent to the ground server whenever a communication window is available and updated asynchronously. To mitigate accuracy degradation caused by stale updates and non-IID data, a proximal regularization term is incorporated into the loss function. Further, a technique to split the model into multiple sections such that they can be transmitted by the satellites within the short contact window under tight transmission energy budget is proposed. Simulation results demonstrate that this approach significantly reduces training time while improving robustness, reducing transmission failures, and maintaining high accuracy.

Topic 2: Real-time Rendering for XR
Presenter: Anish Vipperla

Abstract: Real-time rendering for Extended Reality (XR) experiences must deliver high frame rate and resolution while maintaining low frame noise and user-side latency. XR devices operate under strict memory, compute, thermal, and energy constraints. Hence, real-time ray tracing at the edge in a latency-sensitive, energy and resource-constrained environment is challenging. We propose a cloud-based real-time rendering solution for XR applications which offload key components of the rendering pipeline to the cloud and perform post-processing tasks on the headset. By rendering at low frame rate, resolution, and samples per pixel on the cloud, the rendering time for complex graphics-intensive scenes and 3D environments can be reduced significantly. To enhance the quality of the low-resolution noisy frames, post-processing tasks such as Frame Denoising, Super-Resolution, and Frame Interpolation can be employed on the headset. In this presentation, we present our initial work on implementing MonteCarlo based denoising algorithms for resource-constrained XR devices.


 

April 22, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: ABKI: Ambient Beamforming-Based Keystroke Inference
Presenter: Mingze Pan, Ph.D. student

Abstract: We present ABKI (Ambient Beamforming-based Keystroke Inference), the first system that performs general keystroke recognition on physical keyboards using unencrypted Beamforming Feedback Information (BFI) passively sniffed from ambient Wi-Fi devices. Unlike prior CSI-based attacks that require specialized hardware, privileged access, or close-range transceivers, ABKI exploits ubiquitous BFI-capable devices to mount stealthy keystroke inference even over long distances. ABKI introduces a distance-based optimization that improves the uniqueness rate of training-free inference by 2.1x, and a Transformer-based neural network that enables accurate recognition of both letter and non-letter keys. In evaluation, ABKI achieves letter-key inference accuracy up to 95.5%, non-letter key accuracy up to 91.3%, and word recovery accuracy up to 95.2%, significantly outperforming existing work in both inference accuracy and generability. These results show that ABKI enables practical, stealthy keystroke inference in realistic public settings, exposing a powerful new vector for privacy leakage.


 

April 8, 2026

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Risk-Averse Task Offloading for Tactical Edge Computing: A Reinforcement Learning Approach
Presenter: Xuanli Lin, Ph.D. student

Abstract: Tactical edge networks must support mission-critical task offloading under unreliable communications. In such environments, rare but severe failures can cause substantial degradation in task utility, making tail-risk control essential for robust decision making. To address this challenge, we incorporate Conditional Value-at-Risk (CVaR) to explicitly capture catastrophic performance loss under stochastic communication failures. In our last talk, we introduced a distributional dynamic programming (DP) approach that propagates the full loss distribution across decision stages. This provides an exact baseline for risk-aware offloading, but it is inherently memory-hungry because the recursion must carry distributional information rather than only scalar values.

To improve scalability, we further reformulate the original problem as a constrained Markov decision process (CMDP) with a CVaR constraint on cumulative loss and then derive a reinforcement learning (RL) approach based on a Lagrangian relaxation. Since the resulting penalized return is defined at the trajectory level, whereas standard actor-critic methods typically operate on stepwise transitions, we construct shaped per-step rewards from sampled episodes and update the offloading policy using proximal policy optimization (PPO). Preliminary simulation results show that the proposed RL approach achieves comparable risk-averse performance with better scalability than the exact distributional DP baseline.


 

April 1, 2026

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Two-tiered UCB Restless Bandit Approach for Scheduling in Unknown Wireless Networks
Presenter: Tasmeen Zaman Ornee, postdoc researcher

Abstract: Next-generation wireless networks increasingly rely on information sharing to extend coverage and enhance system throughput, where the Base Station (BS) must coordinate transmissions to collaborative user groups. This challenge is increased when both the Markovian channel fading and the intra-group sharing dynamics are completely unknown to the BS. In this paper, we formulate this scheduling problem as a multi-action Restless Multi-Armed Bandit (RMAB). To address this, we first establish a sufficient statistic based on the last observed Channel State Information (CSI) and its Age of Channel State Information (AoCSI) to reduce the state space. Next, we prove that the inter-group and intra-group decisions can be decoupled without loss of global optimality. Building upon this theoretical foundation, we develop an Online Maximum Group Index First (Online-MGIF) policy. This algorithm executes a double Upper Confidence Bound (UCB) framework to simultaneously learn the unknown intra-group sharing probabilities and the inter-group channel transitions. In addition, the proposed Online-MGIF policy achieves sublinear regret with respect to the number of episodes.


 

March 25, 2026

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Nonparametric Learning of Sensor Systematic Errors in MIMO Systems
Presenter: Yongho Seo, PhD student

Abstract: Sensor systematic errors, such as bias offset, drift, and physical misalignment, can result in significant and persistent performance degradation. This is particularly important in MIMO systems for mmWave propagation since small wavelengths and large number of antenna elements make systems highly sensitive to phase mismatches and environmental changes. Recent methods to compensate for such errors are based on estimating fixed parameters that model sensor bias. However, the errors are sensor-specific and can vary nonlinearly with time. We propose to sequentially learn, and compensate for, sensor systematic errors using Gaussian process (GP) nonparametric modeling, which is a powerful machine learning method used for nonlinear regression and to make predictions with a measure of uncertainty. Within the GP framework, we model the errors as unknown additive stochastic functions of spatiotemporal features to account for offsets in time, range and bearing. Using simulations, we demonstrate the improved tracking performance when the GP learns the error functions at the current time step using training observations from previous time steps. We also propose a new data association algorithm that integrates the GP with belief propagation, an iterative message-passing algorithm used to perform probabilistic inference on graphical models.


 

March 18, 2026

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic 1: Blind Wideband Receiver System for LFM Chirp Detection

Abstract: Radio-frequency (RF) spectrum is increasingly congested with emissions from radar, communication, and sensing systems across wide bands, demanding advanced spectral awareness for interference management and passive monitoring. This paper presents a blind wideband receiver system for linear frequency-modulated (LFM) chirp detection and estimation without prior knowledge of the waveform. The proposed complex ratio cyclic-difference (CRCD) method exploits quadratic phase coherence, enabling detection without reference and closed-form estimation of chirp rate and center frequency. A Neyman–Pearson framework ensures controlled false-alarm rates, while multi-channel averaging improves robustness under low SNR. The algorithm is validated through MATLAB simulations and executed on an FPGA-based emulation of a heterogeneous system-on-chip for real-time block-wise processing. The results demonstrate accurate parameter estimation and reliable detection, confirming the feasibility of CRCD-based blind chirp sensing for real-time passive spectrum monitoring.

Topic 2: Joint Radar and OFDM/5G Receiver with Slow-Time Clutter Mitigation

Abstract: This project studies Integrated Sensing and Communications (ISAC) using a joint OFDM + LFM waveform where a base station acts as both a communication receiver (gNB) and a monostatic radar. In such systems, moving targets produce non-zero Doppler shifts, while static objects in the environment generate strong zero-Doppler clutter, which can mask slow targets and degrade radar detection. The project develops a receiver framework that processes the signal for both communication decoding and radar sensing, generating range-doppler information from the received waveform. A slow-time temporal mitigation approach is explored to suppress static clutter and improve target detectability. Currently, the communication and radar processing chains are implemented and validated separately, and the next step is to integrate them into a unified ISAC receiver and evaluate performance in both simulation and hardware experiments.

Topic 3: Distributed Coherence in Future Generation Wireless Applications

Abstract: Distributed coherence, the act of using spatially separated and independent sensors cooperatively, provides substantial benefits over single sensor platforms in wireless applications. This discussion highlights some of the future applications in both communications and radar, the difficulty in constructing a cooperative system, and the iterative progress we have made in building distributed coherent systems. Future work is proposed to further advance the technology.


 

March 11, 2026

Topic: Heterogeneous and Low Probability of Detection Networking
Presenter: Fikadu Dagefu, DEVCOM Army Research Laboratory

Abstract: State-of-the-art wireless communication technologies often lack resilience, particularly in physically complex and infrastructure-poor environments. Another key challenge stems from the growing need to contend with sophisticated adversarial agents and the threat they pose to both civilian and military networks. This necessitates novel wireless communication capabilities that are agile and less detectable. In this talk, I will discuss our recent research efforts focused on tackling these challenges. In the first part, I will provide an overview of our recent intramural research on the Extremely Heterogeneous Networks (XHN) program. We envision a tightly integrated, intelligent networking capability that exploits diverse and complementary wireless communication modalities through intelligent dynamic adaptation to support multiple simultaneous requirements. Specific results to be highlighted include decentralized learning based covert routing strategies, which establish routes by selecting the optimal set of communication modalities at each relay node. Our custom ns-3-based XHN modeling framework will also be briefly introduced. In the second part of the talk, I will describe our work on Low Probability of Detection (LPD) techniques that leverage novel coding and physical layer technologies, including reconfigurable intelligent surfaces (RIS) and polarization and pattern diversity antennas.


 

February 25, 2026

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: TrustSense: Structure-Adaptive Trust Inference for Resilient Cooperative IoT Sensing
Presenter: Ruobing Li, PhD student

Abstract: Cooperative IoT sensing systems operate in dynamic and often adversarial environments, where noise, sensor drift, malicious manipulation, and evolving device participation can distort inter-device relationships and degrade sensing reliability. Existing approaches typically rely on fixed or statically learned graph structures and provide device-level trust assessments, limiting their ability to adapt to structural changes or evaluate the trustworthiness of individual sensing reports in real time. We propose TrustSense, a structure-adaptive trust inference framework for resilient cooperative IoT sensing. TrustSense employs an inductive graph-based model that performs layer-wise edge mask refinement during training to suppress unreliable interactions and learn a stable cooperation subgraph. A forced anchor attention mechanism preserves the influence of historically consistent devices, stabilizing message propagation under noise and adversarial manipulation. At runtime, TrustSense combines lightweight historical credit initialization with per-report trust inference, enabling robust evaluation of newly joined and intermittently active devices. Extensive experiments on synthetic and real-world datasets demonstrate that TrustSense consistently outperforms representative baselines in accuracy, robustness, and scalability, maintaining reliable trust inference under diverse attack strategies and dynamic sensing conditions.


 

February 18, 2026

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Hierarchical Federated Learning

Abstract: Hierarchical Federated Learning (FL) is a scalable solution for distributed training in a device–edge–cloud hierarchy, where device–edge clusters run split federated learning and a remote cloud server aggregates models across clusters. While this architecture improves privacy and scalability, it faces critical challenges due to intra-cluster latency from frequent activation/gradient exchanges, and inter-cluster latency from long-distance edge–cloud synchronization. In addition, non-IID data across clusters can lead to model divergence. To address these challenges, we propose an end-to-end time-efficient training framework. At the intra-cluster level, we introduce a pipelined scheme that enables clients to overlap computation with communication and thereby hide the communication latency. At the inter-cluster level, we adopt asynchronous aggregation to remove global synchronization related problems. To mitigate the accuracy degradation from model staleness and non-IID data, we incorporate logit calibration and a proximal regularization term into the loss function. Simulation results demonstrate that our approach significantly reduces end-to-end training time while maintaining accuracy comparable to synchronous baselines.

Topic 2: RIS-aided Indoor Device-Free Sensing for Static Object Localization

Abstract: We consider the problem of indoor static target detection/localization for the scenario where the transmitter (TX) and target are inside the room and the receiver (RX) is outside. Most Wi-Fi sensing work focuses on dynamic targets, and localization typically requires multiple TX–RX links. For detection, we propose a RIS-aided approach that uses only one TX–RX pair. We sweep the RIS reflection angle and compare the received signal strength with and without a target. For localization, we use two RISs and estimate the target position from the intersection of the predicted angle regions. Preliminary set of experiments show that these methods work for a sector of radius 2.4m and angle of 60 degrees. Our next steps are to incorporate more advanced signal processing techniques for increasing the localization range and also train an MLP to classify reflection angles for improved static target detection and localization.


 

December 10, 2025

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: NeighborWatch: Collaborative Detection of Malicious Devices in Overlapping Camera Networks
Presenter: Yu Zhou, PhD student

Abstract: Overlapping Camera Networks (OCNs) play a critical role in applications such as urban surveillance, autonomous driving, and public safety. However, their distributed architecture exposes them to threats from compromised cameras capable of submitting forged images. In this presentation, we present NeighborWatch, the first unified framework to detect malicious nodes in OCNs by leveraging the inherent redundancy provided by overlapping fields of view (OFoV). Our approach comprises three complementary modules: Local Neighborhood Detection (LND), which verifies geometric and photometric consistency across neighboring views; Network-Level Detection (NLD), which propagates trust via iterative credit distribution to detect colluding clusters; and Latent Feature Consistency (LFC), which examines semantic coherence between OFoV and non-OFoV regions within each image. Extensive evaluations on real and simulated datasets demonstrate that NeighborWatch outperforms state-of-the-art image forgery detection techniques, achieving high detection accuracy and robustness across diverse attack strategies and forgery types, all with minimal training overhead.


 

December 3, 2025

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Learning and Dynamic Decision-Making in Time-varying Shared Spectrum Environments
Presenter: Abigail Eastman, PhD student

Abstract: Limited resources and high traffic demand in future generation wireless systems necessitate methodologies to dynamically adapt to time-varying and unknown sensing environments. In our previous presentation, we presented a radar tracking method that jointly adapts the transmit waveform and transmission sub-band while learning the time-varying environment using Q-learning (QL). In this presentation, we generalize the approach to the double QL while-tracking (DQL-WT) algorithm to avoid the minimization bias inherent to QL problems. Using the double QL, we explore action selection methods under sudden changes in communications interference levels to improve tracking accuracy. We also consider tracking under incorrect assumptions on the time-varying environment.

We extend our method to scenarios that balance the joint need of transmit waveform selection and sub-band utilization as the number of available sub-bands varies over time. In particular, we propose a double QL based method that dynamically switches the underlying Markov model to account for time-varying changes in the state-action space. The method has the ability to transfer the learning acquired from the prior model to a new state-space representation without loss of information or performance. This allows for real-time reactivity to various unpredictable situations such as adversary jamming, providing active anti-jamming capabilities, without any prior information, in unknown complex environments. The method is also important when sub-bands become inaccessible due to planned emergency service or cooperation with communication users sharing the same spectrum.

Using the DQL-WT method, the radar system learns to avoid sub-bands with high congestion. Indirectly, this may reduce the interference that the radar causes to the co-existing communications system. In order to more directly improve the communications system performance, we are currently extending our approach to adopt waveform selection for both systems while learning the more challenging environment with an increased cost.


 

September 24, 2025

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Risk-Averse Task Offloading for Tactical Edge Computing
Presenter: Dr. Zhaofeng Zhang, postdoc researcher

Abstract: The explosive growth of IoT and mission-critical applications over tactical edge networks demands task offloading strategies that are not only resource-efficient but also resilient to stochastic communication failures. A natural formulation is to maximize the expected total rewards across all users, which provides strong average-case performance and is computationally tractable using dynamic programming (DP). This expectation-based approach serves as a valuable baseline for many network scenarios. However, in adversarial or failure-prone environments, average performance alone may not be sufficient. Rare but severe communication failures can lead to significant performance drops, which are particularly undesirable in mission-critical settings. To address this challenge, we propose a risk-averse offloading framework that explicitly models probabilistic link failures and captures the associated performance uncertainty. Our approach leverages Conditional Value-at-Risk (CVaR), a well-established risk measure from finance and operations research, to characterize the expected loss in the worst-case scenarios. By maximizing the expected total rewards under CVaR constraints, we can achieve robustness against catastrophic performance drops, even if this requires modest sacrifices in expected reward. The central technical contribution of this work is a distributional DP algorithm that propagates full probability mass functions of loss across decision stages. This enables exact evaluation of CVaR at the final stage. To control complexity, we employ Pareto frontier representations to track only the non-dominated trade-offs between reward and risk, and we prune dominated candidates using first-order stochastic domination, which guarantees non-optimal solutions are discarded. Preliminary simulation results demonstrate the trade-off between expected performance and robustness, highlighting how the CVaR-based method complements expectation-based baselines in providing stronger resilience for future generation tactical networks.


 

September 10, 2025

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic 1: Theory of Mixture-of-Experts in Continual Learning
Presenter: Dr. Hongbo Li, postdoc researcher, Ohio State University

Abstract: Catastrophic forgetting is a major challenge in continual learning (CL), where models lose old knowledge when adapting to new tasks. While Mixture-of-Experts (MoE) models have shown practical success in mitigating this issue, their theoretical underpinnings in CL remain unexplored. This work presents the first theoretical analysis of MoE's performance in CL, using an overparameterized linear regression framework. We demonstrate that MoE outperforms a single expert by enabling its specialized experts and intelligent router to distribute tasks effectively. Our analysis reveals that for convergence, the gating network's updates must be terminated after sufficient training. We also show that adding more experts does not always improve performance and can even slow down convergence. We quantify these benefits and trade-offs through explicit expressions for forgetting and generalization error. Finally, we discuss extending our theoretical framework to networking applications like mobile edge computing, where each edge server can be modeled as an expert with delay.

Topic 2: Online Wireless Scheduling for Throughput Maximization of Finite State Markovian Channels with Unknown Statistics
Presenter: Dr. Tasmeen Ornee, postdoc researcher, Ohio State University

Abstract: We consider a wireless scheduling problem in downlink wireless networks with unknown channel statistics. Scheduling performance relies heavily on accurate channel state information (CSI), which is often costly to acquire. In this study, CSI is obtained from ACK/NACK feedback, only after each scheduled transmission. Due to limited channel resources, all users cannot be scheduled for transmission simultaneously. Hence, the most recently observed CSI can be outdated. The traditional approach to solving scheduling problems using outdated CSI is to utilize belief states, which are calculated using the time correlation statistics of channels. However, channel statistics are often unknown; consequently, belief states can be uncountable and this approach becomes infeasible. In this study, we introduce a new sufficient statistic-the latest observed CSI and its Age of Channel State Information (AoCSI), which characterizes the CSI staleness, to make the scheduling decisions. Accordingly, we are able to significantly reduce the state space. We develop an online Maximum Gain First (Online-MGF) policy which achieves sub-linear regret on the number of episodes. Numerical results demonstrate that Online-MGF policy converges to MGF and Whittle index policies with known channel statistics within a very few episodes. In addition, Online-MGF outperforms Maximum AoCSI First (MAF) and random policies.


 

September 3, 2025

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: Subscription-Based Integrated 5G-Radar Tracking System With Dynamic Resource Management Using Reinforcement Learning
Presenters: Yukang Fu and Andrew Herschfelt

Abstract: With radar-as-a-subscriber technology (RAST), radar can be integrated into 5G networks and partially use 5G waveforms for radar applications without extra cost for new waveform design and infrastructure. According to previous works, even with only part of the 5G bandwidth and without using specific allocation in the 5G bandwidth, the radar system can still achieve relatively good detection and estimation performance. However, we still need to improve the joint detection-estimation performance in more complicated tasks, such as tracking, by optimizing the selection of frequency allocations used for radar in a 5G network. In addition, environmental variations during target movement, such as noise, interference, and attenuation, require the selection of radar allocations to adapt. In this paper, we use reinforcement learning to adaptively optimize the allocation of 5G resources to radar users while tracking target movement.


 

August 13, 2025

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Chiplet-based Neural Network Accelerator
Presenters: Prof. Chaitali Chakrabarti and Qifeng Chen

Abstract: We present a multi-chiplet neural network accelerator, where each chiplet consists of a systolic array-based compute fabric, interconnected via a mesh-based on-chip network that utilizes direct signaling. To reduce on-chip activation storage and layer latency, we propose a layer-wise interleaved scheduling strategy that enables overlapped execution of layers. We develop analytical models for compute, storage and interconnect components to evaluate area and timing trade-offs. These models guide the selection of architectural parameters such as array size, number of arrays per chiplet and corresponding layer mappings to optimize throughput or end-to-end latency under die area constraints. We develop lightweight heuristics that deliver near-optimal solutions while avoiding the computational overhead of exhaustive design space exploration. Our heuristic solutions are, on average, within 3.5% difference of those obtained by exhaustive search for characteristic workloads with about 200x lower search time.

Topic 2: RIS-aided through-wall sensing for indoor target detection
Presenters: Prof. Chaitali Chakrabarti and Qifeng Chen

Abstract: We consider the problem of indoor target detection for the scenario where the transmitter (TX) and target are inside the room and the receiver (RX) is outside. While existing work has shown that the wall between the TX and RX helps expand the detection area in the room, there are still some areas that are not covered. We propose use of a Reconfigurable Intelligent Surface (RIS) to increase the sensing coverage. We develop models for RIS-aided through-wall sensing and use SSNR (sensing-signal-to-noise ratio) to quantify the sensing capability. Preliminary simulation results show that the largest sensing area coverage is achieved if the TX is placed on the wall farthest away from the RX, and the RIS are placed on the sidewalls closest to the TX. Our next steps include establishing a firm relationship between the room dimensions and the RIS positions relative to the TX and RX, and conducting experiments with RIS platforms to validate our results.


 

July 30, 2025

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Dynamic Decision Theoretic Learning: Merged Measurements and Dynamic Dependencies for mmWave Radars
Presenter: Yongho Seo, PhD student

Abstract: Although mmWave radars provide high range and velocity resolution, they have limited angular resolution depending on the number of antennas used. Sensor resolution limitations can result in merged measurements that occur when multiple closely spaced targets are perceived as a single target. As signals are more susceptible to attenuation and interference at high frequencies, merged measurements become problematic when tracking multiple targets in dynamic and dense environments with high clutter intensity. One example of such a scenario is urban traffic with many vehicles and pedestrians. We consider the challenging problem of merged measurements when tracking an unknown and time-varying number of targets, with unknown labels and unknown measurement associations. To improve tracking performance by reducing track loss and target mismatch, we extend our nonparametric Bayesian learning and tracking approach to include validation gating and merged measurement joint association. The method also adapts a target lifetime parameter to reduce track fragmentation. The proposed approach scales efficiently as the number of targets increases, and dynamic programming is implemented to meet operational constraints. Simulations demonstrate significant improvements in processing merged measurements and increase tracking accuracy for a varying number of closely-spaced targets with intersecting tracks.


 

July 23, 2025

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Delay-Tolerant Relay-Aided Task Offloading for Tactical Edge Computing
Presenter: Zhaofeng Zhang, Postdoctoral Researcher, Arizona State University

Abstract: The growing deployment of intelligent services over tactical edge networks necessitates resilient and efficient task offloading frameworks that can operate in dynamic and resource-constrained environments. While prior works impose hard deadline constraints on task execution, such rigid assumptions often limit offloading flexibility and degrade system performance. In this work, we propose a delay-tolerant task offloading framework that generalizes existing models by allowing each user to flexibly execute the task after its deadline within its tolerable range. This enables a fundamental trade-off between delay and system efficiency, as larger execution windows can reduce total time slot consumption and improve overall resource utilization. Built upon a multi-hop communication architecture that enables tasks to be executed not only at edge servers but also at relay nodes and local devices, the proposed framework jointly optimizes task placement, resource allocation, and TDMA-based time slot scheduling. The resulting problem is formulated as a nonlinear integer program aimed at maximizing the total reward under connectivity and computing resource constraints. We develop an optimal dynamic programming algorithm that leverages a feasibility-aware network graph and identifies the slot-minimizing offloading paths and execution durations. Additionally, scalable heuristics are proposed via scaling and rounding techniques. Experimental evaluations demonstrate that the delay-tolerant approach consistently outperforms hard-deadline models in terms of total reward, while offering flexible control over execution delay and maintaining high system efficiency across diverse test cases.


 

July 16, 2025

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Wireless Scheduling for Throughput Maximization with Unknown Channel Statistic
Presenter: Tasmeen Ornee, Postdoctoral Researcher, Ohio State University

Abstract: In this study, we consider a wireless scheduling problem where a Base Station (BS) sends time-sensitive information to multiple users without knowing the accurate channel state information (CSI). The traditional approach to solve scheduling problem with known CSI is to utilize belief states. However, this approach is infeasible in unknown CSI environment due to curse of dimensionality. On the other hand, because the communication resources are limited, the received information is not always fresh. One performance metric to characterize data freshness is the Age of Information (AoI). We show that the latest received CSI and its AoI is a sufficient statistic of the history to make the scheduling decisions and by this, we are able to significantly reduce the state space. The formulated problem is a Restless Multi-armed Bandit (RMAB). Our goal is to develop an online scheduling algorithm that maximized the expected discounted throughput of all users over an infinite time-horizon while satisfying a channel resource constraint. To that end, we develop an online Maximum Gain First (MGF) policy which does not need to satisfy any indexability condition. Our algorithm achieves sub-linear regret on the number of episodes.


 

July 9, 2025

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: RIS-Key: Scalable and Secure Physical-Layer Key Generation via Optimized Reconfigurable Intelligent Surfaces
Presenters: Yanchao Zhang & Yizhou Wang

Abstract: Physical-Layer Key Generation (PLKG) enables two wireless devices to independently derive identical cryptographic keys by leveraging the reciprocal and correlated variations of their shared wireless channel. In mission-critical, large-scale IoT networks that lack a trusted central authority, secure device-to-device communication must rely on such decentralized key establishment. However, traditional PLKG schemes often fail in static or slowly varying environments due to insufficient channel randomness. To overcome this limitation, Reconfigurable Intelligent Surfaces (RIS) offer a promising way to induce fast, controllable channel variations. Yet, existing RIS-assisted PLKG methods have not explored optimizing RIS configurations, resulting in suboptimal key generation performance in large-scale networks. In this talk, I will introduce RIS-Key—our latest RIS-assisted PLKG framework designed for fast, secure, and scalable pairwise key generation in dense IoT deployments. RIS-Key adaptively selects RIS codewords based on real-time channel conditions and device groupings, aiming to maximize key generation rate while minimizing key disagreement. I will present the design principles of RIS-Key, highlight its performance improvements over baseline methods based on preliminary experiments, and discuss open research challenges moving forward.


 

July 2, 2025

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: Adaptive Waveform Co-design for Integrated Sensing and Communications (ISAC)
Presenters: Dan Bliss, Andrew Herschfelt, and Shammi A. Doly

Abstract: This study explores the potential of advanced adaptive waveform co-design strategies for Integrated Sensing and Communications (ISAC) systems, targeting the future demands of 6G and next-generation wireless networks. The goal of this research is to develop ISAC algorithms that co-design traditional radio detection and ranging (radar) waveforms to enable accurate sensing and high-quality communications simultaneously, through the seamless integration of macchine learing (ML) within constrained hardware and spectral resources. Moreover, the feasibility of the developed ISAC algorithms was evaluated through various Hardware-in-the-Loop (HWIL)-over-the- air (OTA) experimental testbeds, including WISCANet and GNU Radio. The waveform co-design problem is formulated across different multi-access scenarios and environmental conditions, involving both single and multiple ISAC nodes operating under the “Network as Sensors (NaS)” paradigm—enabling simultaneous communication and sensing through an optimized radar waveform. A decision-theoretic framework is employed to jointly optimize performance in spectrally shared environments, with the initial formulation based on a partially observable Markov decision process (POMDP) for single-node configurations. To address the resulting optimization problem, two approximate dynamic programming methods— nominal belief-state optimization (NBO) and random-sampling multipath hypothesis propagation (RS-MHP)—are extended and systematically evaluated. In addition to traditional linear frequency modulation (LFM) chirps, we explore the opportunities of using a novel Non-linear frequency modulation (NLFM) chirp co-design methodology based on the principle of stationary phase (PSP) concept. This approach yields a closed-form solution for the spectrum of the NLFM chirp with a parametric polynomial phase in terms of signum function, highlighting its performance characteristics on the Cramer-Rao lower bound (CRLB) in the ISAC scenario.


 

June 4, 2025

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Addressing Communication and Computation heterogeneity in Split Federated Learning schemes
Presenters: Prof. Chaitali Chakrabarti and students

Abstract: Split Federated Learning (SFL) is a privacy-preserving distributed machine learning framework suitable for resource-constrained clients. It splits the neural network model between client and server such that most of the computations are offloaded to the server, without sharing private client raw data to server. In our previous work (TEE-SFL), we addressed the heterogeneity in the channel conditions of different clients by applying different degrees of compression on the data to be communicated and also adding an auxiliary network at the client-end to facilitate local training during times when the clients cannot communicate to the server. This talk will focus on our recent work in also addressing the heterogeneity in clients’ computation capabilities. We address this gap by compressing the model at the client end through use of singular value decomposition (SVD). We show that a combination of SVD and bottleneck layers can help reduce the training time with only a small drop in accuracy compared to TEE-SFL.

Topic 2: Customized Design of Chiplet-based Neural Network Accelerators
Presenters: Prof. Chaitali Chakrabarti and students

Abstract: The increasing size and complexity of Deep Neural Networks (DNNs) has intensified the demand for scalable and high performance hardware accelerators. Multi-chiplet accelerators have emerged as a promising solution to meet these demands. However, designing and optimizing such systems for high throughput or low end-to-end latency with die area constraints is very challenging. Our starting point is a multi-chiplet architecture, where each chiplet consists of a systolic array-based compute fabric, interconnected via a mesh-based on-chip network that utilizes direct signaling. We first develop analytical area and timing models for compute and storage. To reduce on-chip activation storage and latency, we propose a layer-wise interleaved scheduling strategy that enables overlapped execution of DNN layers. For a given die area, we derive architectural parameters, such as systolic array size, number of arrays in a chiplet and the corresponding layer mappings to achieve maximum throughput or minimum end-to-end latency. We show that for both VGG-16 and ResNet-18, the layer-wise interleaved scheduling strategy reduces the memory footprint significantly while significantly improving the timing performance.


 

May 28, 2025

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: RIS-CLA: Reviving CSI-Based Continuous Location Authentication with Reconfigurable Intelligent Surfaces
Presenters: Yanchao Zhang & Yan Zhang

Abstract: Continuous Location Authentication (CLA) enables continuous and autonomous verification of a networked device’s physical location, preventing unauthorized movement or removal. When tied to a specific user (e.g., post-login), verifying the device’s location also affirms the user’s physical presence. CLA is particularly critical for securing access to sensitive resources in mission-critical facilities, hospitals, and corporate environments. Wireless Channel State Information (CSI) is a promising enabler for CLA, as it captures fine-grained physical-layer characteristics that are inherently tied to a device’s location and surroundings. However, recent studies have revealed that state-of-the-art CSI-CLA systems are vulnerable to CSI-guessing attacks, in which adversaries infer a victim’s CSI-based location profile and reproduce it by encoding Wi-Fi signals from another device, effectively spoofing the location as perceived by the verifier. We propose RIS-CLA, a system that leverages Reconfigurable Intelligent Surfaces (RISs) to enable secure and autonomous CSI-based CLA, resilient to CSI-guessing attacks. The key idea behind RIS-CLA is to induce rapid, controlled, and unpredictable wireless channel variations by periodically altering RIS configurations during the authentication session. This process generates time-varying CSI-location profiles that are verifiable by the legitimate party but infeasible to predict or emulate by attackers. The effectiveness of RIS-CLA is validated through real-world experiments using our in-lab RIS prototypes.


 

May 14, 2025

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Multi-Hop Relay-Aided Task Offloading for Tactical Edge Computing
Presenter: Dr. Zhaofeng Zhang (postdoc researcher)

Abstract: Recent years have witnessed explosive growth in the deployment of IoT devices over tactical edge networks, which demand flexible task offloading strategies to accommodate limited connectivity and computing resources. While many existing works assume direct access to edge servers, such assumptions often break down in adversarial and dynamic environments. This work introduces a novel edge task offloading framework that enables multi-hop communication and enables task execution not only at edge servers but also at relay nodes and local devices. We formulate the problem of joint task offloading, resource allocation, and time frame/slot assignment, under a time-division multiple access (TDMA) uplink scheme, as a nonlinear integer program to maximize total rewards, subject to various communication and resource constraints. To solve it optimally, we first construct a feasibility-aware directed graph that captures valid communication links based on signal-to-noise ratio (SNR) constraints, and then design a dynamic programming (DP) algorithm that integrates slot-minimizing offloading path selection with the assignment of the largest feasible number of time frames for each user to reduce total time slots consumption without violating task completion deadlines. To improve scalability, efficient heuristics are developed using scaling and rounding techniques. Extensive experiments demonstrate that our proposed DP algorithm achieves optimal task offloading and resource allocation, while our heuristic algorithms offer a computationally efficient alternative with competitive performance.


 

May 7, 2025

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Online Fair Reinforcement Learning for Restless Bandits
Presenter: Tasmeen Zaman Ornee, PhD student

Abstract: Restless Multi-armed Bandit (RMAB) framework is a popular approach to tackle resource allocation problems in multi-agent systems where the dynamics of each arm are modeled as a Markov Decision process (MDP). In this study, we aim to find a Reinforcement Learning (RL) algorithm that maximizes the sum of time-averaged expected reward of all agents over a finite time-horizon with unknown system dynamics. However, traditional RMAB approaches often ignore fairness among the allocated resources of all agents: an agent with a higher reward will get resources frequently and agent with a lower reward may wait for a long time. We address this limitation by considering a fairness constraint. To this end, we develop a novel Maximum Gain First (MGF) policy that does not need to satisfy any indexability condition. The developed algorithm is computationally efficient and more simplified than existing studies. Our algorithm achieves sublinear regret, and the simulation results exhibit better performance compared to existing methods.


 

April 30, 2025

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Feature-based learning for computation-efficient estimation in high frequency time-varying spectrum sharing environments (In collaboration with Chaitali Chakrabarti)
Presenter: Pinjia Zhang, PhD student

Abstract: Sharing spectrum between radar and communication systems in high frequency environments is essential to support the rapid advances in wireless technology with only a limited amount of resources. We consider the problem of estimating the position of a target when the radar is sharing spectrum with multiple communication users and thus faces high levels of interference. We propose an adaptive estimation method that aims to maintain high estimation accuracy while reducing the need for computationally intensive estimation algorithms for real time processing. The adaptive method selects between two estimation algorithms using a threshold that depends on the signal-to-interference-plus-noise ratio (SINR). The first method is the asymptotically efficient maximum-likelihood estimator. Although fast to implement, this estimator approaches optimal performance in maximizing the measurement likelihood only at high SINRs. The second method is computationally intensive to implement but maintains high accuracy at low SINR. This method is our extension of a learning-based approach that combines time-frequency feature extraction with Gaussian-mixture modeling. Using simulations under high frequency spectrum-sharing conditions, we demonstrate the computation-efficiency and estimation performance of the proposed adaptive method, when compared to other methods, under time-varying SINR conditions.


 

April 23, 2025

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Presenter: Yukang Fu

Abstract 1: Spectral resources have been running short, and it is inevitable for radar and communication users to share spectrum, which can cause performance loss. While there are several methods to potentially address this problem, integrating radar systems into existing cellular networks offers a low-cost solution. Radar users can subscribe as multiple communication users in the communication network to utilize part of the communication bandwidth as radar waveform. In this paper, we simulate the subscription-based joint 5G-radar system in MATLAB and discuss the radar performance and sidelobe level distribution using different waveforms generated from different spectral allocations in 5G networks, including using randomly selected frequency bands, edges and center of the spectrum, in order to find a balance between cost and performance.

Abstract 2 (if time permits): Bit-domain data framing is a crucial component in communication systems, where operations such as scrambling, error corrections, and interleaving are conducted. It ensures the efficiency, reliability, and adaptability of data transmissions. Specifically, the demands on flexible optical communications have been raised by the requirements for scalability and energy efficiency in modern communication systems. Therefore, a reconfigurable bit-domain framing architecture is necessary. In this paper, an efficient bit-domain framing architecture for flexible optical communications will be introduced. The proposed design would allow for reconfiguration, enabling it to support different standards and potential protocols that haven't been deployed yet. By beginning with the OpenZR+ and SDA OCT v3.0 standards, we presented a first design for this purpose. Additionally, the results from the implementations with GF-12nm FinFET technology are presented to demonstrate the performance. Running at 1GHz, it is able to achieve a throughput of around 111 Gbps in OpenZR+ mode. In SDA OCT v3.0 mode, the proposed design can meet a throughput from 10 Gbps to 57 Gbps along with different payload coding rates.


 

April 2, 2025

Thrust 7: FutureG AR/VR Applications
Lead: Robert LiKamWa
 

Topic 1: AutoCalNet, a real-time calibration system for mobile multi-camera rigs
Presenter: Flemming Laursen

Abstract: We will present our ongoing research on AutoCalNet, a real-time calibration system for mobile multi-camera rigs. The system eliminates the need for manual checkerboard calibration and allows cameras to move without disrupting the setup. While multi-camera systems are widely used to generate 3D point clouds for telepresence, healthcare, and immersive media, their scalability is often limited by the complexity and fragility of calibration. AutoCalNet addresses this by relying solely on video feeds and modern vision transformers to continuously estimate both intrinsic and extrinsic parameters, enabling flexible and reconfigurable rigs. We will also introduce the concept of an enriched 3D representation that serves as a shared context across multiple neural networks, supporting spatial reasoning and calibration tasks. Finally, we will discuss the challenges and opportunities involved in developing protocols to improve accuracy and coordinate camera behavior, and how this approach opens new possibilities for building robust, adaptive multi-camera systems.

Topic 2: Game engine based adaptive volumetric streaming using 5G and MEC
Presenter: Krutik Pandya

Abstract: Volumetric Media is the next generation of media, in which we deal with rich 3D data, including point clouds, meshes, gaussian splats, neural rendering fields, and other 3D scenes. The adoption of this media has been a challenge due to its massive size and streaming becomes difficult as we scale the data. In this presentation, we are proposing a method that can allow us to stream and view volumetric data in real time using 5G and Multi-Access-Edge infrastructure. We utilize game engines to render and optimize the media, and virtual cameras to stream the volumetric data. Our goal is to share the volumetric data to multiple independent viewers using adaptive bitrate streaming. We focus on whether people are in groups or separate to offload the data based on a consensus of devices.


 

March 26, 2025

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: TEE-SFL: A time and energy-efficient solution for addressing communication heterogeneity in Split Federated Learning schemes
Presenters: Prof. Chaitali Chakrabarti and students

Abstract: Split Federated Learning (SFL) is a privacy-preserving distributed machine learning framework suitable for resource-constrained clients. It splits the neural network model between client and server such that most of the computations are offloaded to the server, without sharing private client raw data to server. One drawback of the system is the large communication cost caused by frequent activations/gradients exchanges between client and server. Previous SFL works have focused on reducing the communication cost of homogeneous clients without considering communication heterogeneity due to variations in channel conditions. In this paper, we focus on a more realistic scenario, where clients have to communicate under different channel conditions and also handle non-IID data. We address the heterogeneity in the channel conditions of different clients by applying different degrees of compression on the data to be communicated and also adding an auxiliary network at the client-end to facilitate local training during times when the clients cannot communicate to the server. The experimental results on VGG11 and ResNet18 on CIFAR10 show that compared with other competing methods, our method significantly reduces the training time and energy consumption with very little accuracy drop.

Topic 2: Low-cost and scalable zone-specific CSI feedback method
Presenters: Prof. Chaitali Chakrabarti and students

Abstract: Zone-specific CSI feedback schemes have been shown to improve CSI encoding and decoding performance. However, these methods come with additional computational load and model training overhead. In this talk, we present our approach to achieving low-cost and yet scalable zone-specific CSI feedback. We show that by employing a pre-trained encoder model shared across all zones and training zone-specific decoders, we can achieve performance comparable to the 1-zone baseline while reducing the memory footprint and computational cost by 4x. We also evaluate the scalability of the proposed method for multiple city scenarios. We show that directly using the pre-trained encoder model leads to performance degradation, but partially tuning the encoder can help mitigate part of the loss. We use independent auto-encoder models to learn features from line-of-sight (LOS) and non-line-of-sight (NLOS) paths. The LOS encoder is pre-trained and reused across different zones, while the NLOS encoder is tuned with samples from each zone. We evaluate the separable encoder design using both the Boston downtown scenario (seen by the model during pre-training) and the other six city scenarios (unseen by the model during pre-training) for different ratios of LOS and NLOS encoder sizes. In the Boston downtown evaluation, the results show that when LOS and NLOS are of equal sizes, the method achieves the best CSI reconstruction performance, with 37.5% lower memory footprint and 75% lower computation cost compared to the 1-zone baseline. For unseen city scenarios, the separable encoder design achieves better CSI reconstruction performance while saving up to 50% of tuning cost by using less trainable parameters compared to the method where the entire model is retrained for the new scenario. To further improve the performance of unseen city scenarios, we plan to implement the zone-specific method in these cities using the clustering-based floorplan generating algorithm.


 

March 19, 2025

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: Physical Backdoor Attacks against mmWave-based Human Activity Recognition
Presenter: Yanchao Zhang

Abstract: Human Activity Recognition (HAR) using wireless signals like mmWave technology has promising applications in numerous scenarios, including monitoring and surveillance, healthcare, and smart home. Wireless HAR is non-intrusive and can operate in situations where traditional sensors or cameras may fail. However, these systems also introduce new attack surfaces alongside their benefits. Existing security research on wireless HAR primarily focuses on the vulnerabilities of the AI models used by these systems, without addressing the challenges of physically implementing these attacks in real-world scenarios. In this paper, we present the first physical backdoor attacks for mmWave-based HAR systems, manipulating physical signals to deceive the systems into producing targeted outputs. Utilizing passive metal reflectors and optimized attacking strategies, our attack is efficient, stealthy, and easy to implement. Tailored experiments on a mmWave HAR prototype demonstrate the high effectiveness of the proposed attacks.


 

March 12, 2025

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Task Offloading and Resource Allocation for Relay-Aided Edge Computing
Presenter: Dr. Zhaofeng Zhang

Abstract: Recent years have witnessed an explosive growth in the deployment of IoT devices over tactical edge networks. Given the limited computing resources available at local edge devices, an urgent challenge arises: “how can we design efficient task offloading strategies and resource allocation plans while satisfying the computing resource constraints of servers and the communication constraints between devices and servers?” In our previous work, we addressed this challenge under a single-hop communication setting. In this talk, we extend our framework to the multi-hop case, introducing a more general task offloading paradigm based on a two-hop relaying strategy. Here, relays not only facilitate communication with distant edge servers but also provide moderate computing capabilities for offloaded tasks. Specifically, we formulate the task offloading and resource allocation problem as a multiple multi-dimensional knapsack problem and develop a dynamic programming (DP) solution with pseudo-polynomial time complexity, assuming a bounded number of servers and relays. To further enhance computational efficiency, we incorporate rounding and scaling techniques to reduce the time complexity of DP. Through numerical evaluations, we demonstrate that our proposed exact algorithms achieve optimal task offloading and resource allocation, while our heuristic algorithms offer a computationally efficient alternative with competitive performance.


 

March 5, 2025

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic 1: Designing Provably Efficient Reinforcement Learning in Partially Observable Markov Games
Presenter: Prof. Ming Shi (University of Buffalo)

Abstract: We study multi-agent reinforcement learning under partial observability, where each agent has access only to its own observations and actions, leading to incomplete knowledge of the system’s state. A key challenge is the absence of true state information and joint action policy. For example, in wireless channel scheduling, users often lack full knowledge of channel conditions and other users’ actions due to energy and security constraints. This raises a fundamental question: How much online state information is sufficient for tractability and Nash equilibrium performance? In this talk, I will present the problem setting, key challenges, and the implications of designing provably efficient solutions for this open problem.

Topic 2: Online Learning for Optimizing AoI-Energy Tradeoff under Unknown Channel Statistics
Presenter: Dr. Mohamed Abd-Elmagid (Ohio State)

Abstract: We consider a real-time monitoring system where a source node (with energy limitations) aims to keep the information status at a destination node as fresh as possible by scheduling status update transmissions over a set of channels. The freshness of information at the destination node is measured in terms of the Age of Information (AoI) metric. In this setting, a natural tradeoff exists between the transmission cost (or equivalently, energy consumption) of the source and the achievable AoI performance at the destination. This tradeoff has been optimized in the existing literature under the assumption of having a complete knowledge of the channel statistics. In this work, we develop an online learning-based algorithm with finite-time guarantees that optimizes this tradeoff in the practical scenario of having non-linear age function and unknown channel statistics. In particular, when the channel statistics are known, the optimal scheduling policy is first proven to have a threshold-based structure with respect to the value of AoI (i.e., it is optimal to drop updates when the AoI value is below some threshold). This key insight was then utilized to develop the proposed learning algorithm that surprisingly achieves an order-optimal regret (i.e., O(1)) with respect to the time horizon length.


 

February 26, 2025

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic 1: Dynamic tracking in high clutter and interference from sharing spectrum
Presenter: Yongho Seo, PhD student

Abstract: We present our latest results on tracking multiple targets with an unknown number of radar measurements that are corrupted by environmental clutter and interference from wireless systems that sharie spectrum access. Specifically, we adapt our Dirichlet process mixture model-based measurement-to-target association method with a statistical measure to separate clutter and interference from target measurements. We introduce a statistical approach to correctly manage target labeling when multiple targets are in close proximity to reduce the risk of identity switching. To enhance tracking robustness, we use a track management strategy and an adaptive parameter labeling method to to dynamically prevent track fragmentation. Furthermore, we consider various statistical models of wireless communication interference, including Gaussian and non-Gaussian models, based on the application. We apply a generalized likelihood ratio test to detect the existence of communication interference and estimate its intensity to improve the target-to-measurement association performance.

Topic 2: Q-learning exploration strategies for adaptive waveform and sub-band selection in spectrum sharing environments
Presenter: Abigail Eastman, PhD student

Abstract: We consider the problem of dynamically adapting the transmit waveform and frequency sub-band of a radar system sharing spectrum with a multiuser wireless communications system. As the communications interference varies with the number of users, the environment is assumed unknown and time-varying. Our approach to minimizing both interference and tracking error is based on the Q-learning algorithm; this is a reinforcement learning method where the system learns how to optimally achieve its goal based on its real-time interactions with the environment without knowledge of the environment dynamics. We compare two action selection methods for use in Q-learning while tracking: Epsilon-Greedy and Upper Confidence Bound (UCB). Epsilon-Greedy manages the exploration versus exploitation dilemma by taking random actions an epsilon percentage of the time and exploiting learned knowledge the rest of the time. UCB selects actions based on their uncertainty measure and estimated value. An action’s uncertainty depends on the amount of time and how often it has been selected. Using simulations, we demonstrate that the UCB method increases tracking accuracy and interference mitigation due to the resulting lower average regret and lower percentage error in action selection for both the waveform and sub-band.


 

February 19, 2025

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic: Enabling FutureG with the DASH SoC
Presenter: Dan Bliss

Abstract: We discuss the needs of FutureG in the context of both DoD and commercial systems. We suggest that future systems will employ multiple-function RF (communications, radar, PNT, etc.) which requires flexible capabilities. These future systems drive development of several supporting technologies, including novel joint waveform/algorithmic optimization; flexible RF frontends; distributed-coherent capabilities; and — focus of this presentation — flexible, efficient computational systems. Specifically, we introduce our Domain-Focused Advanced Software-Reconfigurable Heterogeneous (DASH) SoC and software framework. We provide motivation for its architecture, an overview of the enabling software approach, and a discussion of the orders of magnitude benefits of our architecture compared to current approaches. We present the current state of DASH development and challenges to get the DASH SoC into fielded systems.


 

February 12, 2025

Thrust 7: FutureG AR/VR Applications
Lead: Robert LiKamWa
 

Topic 1: Benchmarking and Visualizing Compression Errors in Volumetric Streaming Systems
Presenter: Robert LiKamWa

Abstract: Volumetric streaming is a powerful medium that transmits volumetric data, which primarily includes color and depth information, over a network in real-time. While color data can be effectively compressed using standard video codecs, compressing depth data poses a significant challenge as we need to change the bitrate for transmission over standard video codecs. Streaming depth data using standard video codecs introduces visual artifacts that degrade the quality of volumetric streaming workflows. To address this issue, we propose a demonstration exploring efficient compression techniques for color and depth data that minimize visual artifacts while preserving visual fidelity. Using a physical tabletop game filmed with volumetric cameras, we compare two volumetric streaming video workflows. The first workflow streams and renders the color and depth data directly from the camera in an uncompressed format, which is used as the ground truth. The second workflow encodes and streams the data in a compressed format. The comparison is conducted on a desktop system to evaluate performance for analyzing factors such as data transmission, compression quality, and overall system responsiveness.

Topic 2: AutoCalNet - Continuous Real-Time Calibration of Networked Mobile Volumetric Cameras
Presenter: Robert LiKamWa

Abstract: AutoCalNet enables continuous real-time calibration of mobile 3D cameras by decoupling calibration from content. It leverages a scalable device-edge-cloud network to minimize bandwidth, manage latency, and maintain high precision in calibration data, prioritizing trackable regions and feature points that will facilitate spatiotemporal tracking. This approach provides a flexible, efficient solution for networked camera systems without being constrained by content-specific requirements.


 

January 29, 2025

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Reducing training time of Split Federated Learning
Presenters: Chaitali Chakrabarti and Yan Xiong

Abstract: Split Federated Learning (SFL) is an emerging privacy-preserving decentralized learning scheme ideally suited for resource-constrained clients. SFL splits a machine learning model between client and server such that most of the computations are offloaded to the server. While existing SFL schemes focus on homogeneous clients, a more realistic scenario is when clients are heterogeneous, and the heterogeneity can be with respect to both their computing capability and communication data rate availability. Earlier we had described a method to address differences in communication costs across clients through use of bottleneck layers (BL) that support different degrees of data compression and auxiliary networks which allow for data to be periodically sent to the server. In this work we consider compute-level heterogeneity on top of communication-level heterogeneity. We show that for the case when the communication time is larger (or same order of magnitude) as the computation time, the training time (=computation time + communication time) can be reduced without compromising the accuracy by judicious choice of a combination of BL compression and periodic sending of data. In the extreme case when the computation time is larger than the communication time, the training time can be minimized by additionally adjusting the client-side model size through use of Singular Value Decomposition and other model compression techniques.

Topic 2: Low-cost and scalable zone-specific CSI feedback method
Presenters: Chaitali Chakrabarti and Yan Xiong

Abstract: Zone-specific CSI feedback schemes have been shown to improve CSI encoding and decoding performance. However, these methods come with additional computational load and model training overhead. In this talk, we present our approach to achieving low-cost and yet scalable zone-specific CSI feedback. We show that by employing a pre-trained encoder model shared across all zones and training zone-specific decoders, we can achieve performance comparable to the 1-zone baseline while significantly reducing the memory footprint and computational cost. We also evaluate the scalability of the proposed method for multiple city scenarios. We show that directly using the pre-trained encoder model leads to performance degradation, but partially tuning the encoder can help mitigate part of the loss. Next, we will focus on improving the model tuning efficiency by utilizing a separable auto-encoder design. We will use independent auto-encoder models to learn features from line-of-sight (LOS) and non-line-of-sight.


 

January 22, 2025

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: TagStroke: RFID-Based Keystroke Inference Attacks
Presenters: Jiawei Li and Yanchao Zhang

Abstract: Wireless sensing is expected to play a pivotal role in FutureG networks but also presents vulnerabilities that adversaries can exploit, a growing concern known as adversarial wireless sensing. In this presentation, we will discuss our latest findings on adversarial RFID sensing-based keystroke inference attacks, which pose significant threats to data security and user privacy. Existing methods often exploit typing-induced signals, such as vibrations, acoustics, and visual cues, but often require proximity, line-of-sight, or precise transceiver placement, limiting their practicality. We propose TagStroke, a non-invasive, wireless keystroke inference attack leveraging low-cost COTS RFID systems. TagStroke uses passive UHF RFID tags placed beneath keyboards and a concealed RFID reader to detect typing-induced signal changes, enabling accurate keystroke recognition and semantic recovery. TagStroke faces key challenges such as RFID signal instability, interference from hand and body movements, and limited spatial sensing resolution. To overcome these, it introduces an optimized tag layout, robust signal preprocessing, a hybrid Temporal Convolution-Transformer model for keystroke detection, and a multi-stage recognition framework integrating spatial decoding with large language models (LLMs). We prototype and evaluate TagStroke using a low-cost COTS RFID system with 11 volunteers in diverse settings. TagStroke achieves 97.21% keystroke detection accuracy, 87.24% keystroke recognition accuracy, and a word error rate (WER) of 21.06%, along with high content recovery similarity scores, such as 0.71 at 2 m. These results rival existing WiFi-based attacks while avoiding their constraints, demonstrating TagStroke’s practicality, accuracy, and cost-effectiveness.


 

December 11, 2024

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Online Learning for Optimizing AoI-Energy Tradeoff under Unknown Channel Statistics
Presenter: Mohamed Abd-Elmagid

Abstract: We consider a real-time monitoring system where a source node (with energy limitations) aims to keep the information status at a destination node as fresh as possible by scheduling status update transmissions over a set of channels. The freshness of information at the destination node is measured in terms of the Age of Information (AoI) metric. In this setting, a natural tradeoff exists between the transmission cost (or equivalently, energy consumption) of the source and the achievable AoI performance at the destination. This tradeoff has been optimized in the existing literature under the assumption of having a complete knowledge of the channel statistics. In this work, we develop online learning-based algorithms with finite-time guarantees that optimize this tradeoff in the practical scenario where the channel statistics are unknown to the scheduler. In particular, when the channel statistics are known, the optimal scheduling policy is first proven to have a threshold-based structure with respect to the value of AoI (i.e., it is optimal to drop updates when the AoI value is below some threshold). This key insight was then utilized to develop the proposed learning algorithms that surprisingly achieve an order-optimal regret (i.e., O(1)) with respect to the time horizon length.


 

November 6, 2024

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic 1: Transmit Waveform Design using Reinforcement Learning
Presenter: Antonia Papandreou-Suppappola

Abstract: We consider the dynamic transmit waveform design for a radar system sharing spectrum with a wireless multiuser communications system. The radar sensing environment is unknown and time-varying as the level of communications interference varies with the number of users. We sequentially select the nonlinear phase function of the transmit waveform to minimize tracking MSE and the propagation sub-band to minimize interference. Our approach uses the Q-learning algorithm, which is a reinforcement learning method where the system learns how to optimally achieve its goal based on its real-time interactions with the environment without knowledge of the environment dynamics.

Topic 2: Channel Modeling for Millimeter Wave Propagation
Presenter: Antonia Papandreou-Suppappola

Abstract: An important component of future generation wireless systems is the modeling of the propagation medium at the extremely high frequencies of the millimeter wave (mmWave) and terra Hz (THz) bands. Such models are necessary for continued technological advances to improve sensing and communication performance. Although various channel models have been considered during the last decade, there are still many remaining propagation challenges that need to be addressed. We will present an overview of some of these models, with emphasis on spatial-time-frequency representations and diversity techniques for outdoor high mobility environments.


 

October 9, 2024

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Topic 1: Adaptive FMCW Radar Waveform Optimization for Future-G Networks
Presenters: Dan Bliss and PhD students

Abstract: In this research, we develop a highly flexible real-time FMCW radar system for designing and testing different radar waveforms (e.g., triangular, sawtooth) and radar signal processing algorithms (e.g., stretch processing) without the need for extensive hardware modifications using low-cost SDRs in GNU Radio platform. We demonstrate the previously developed decision-theory based adaptive waveform co-design approach in a single real-time feedback loop for multi-user networks applications in a much simplified mannar using the spectral amplitude masking method. To implement a feedback loop in real-time, an asynchronous message-passing interface is employed in this experiment. We show the benefits of radar waveform optimization, particularly by increasing RMS bandwidth and significantly increasing target detection performance by enhancing range resolution. To demonstrate real-time spectrum optimization, we use the GNU Radio block for real-time spectrum analyzer (RTSA)-like spectrum visualization.

In parallel, we investigate the possible advantages of utilizing the NLFM chirp for FMCW radar waveform optimization. Specifically, we propose a faster unimodular NLFM chirps optimization technique for ISACs applications using the PSP concept to advantageously control the shape of PSD exhibiting substantially reduced ACF sidelobes compared to LFM waveforms. This method is particularly useful in applications requiring fast adaptation to changing conditions, such as in autonomous systems, surveillance, and joint sensing-communications networks.

Topic 2: Subscription-Based Joint 5G-Radar System Design and Analysis
Presenters: Dan Bliss and PhD students

Abstract: Spectral resources have been running short, and it is inevitable for radar and communication users to share spectrum, which can cause performance loss. While there are several methods to potentially address this problem, integrating radar systems into existing cellular networks offers a low-cost solution. Radar users can subscribe as multiple communication users in the communication network to utilize part of the communication bandwidth as radar waveform. In this paper, we simulate the subscription-based joint 5G-radar system in MATLAB and discuss the radar performance and sidelobe level distribution using different waveforms generated from different spectral allocations in 5G networks, including using randomly selected frequency bands, edges and center of the spectrum, in order to find a balance between cost and performance.


 

October 2, 2024

Thrust 7: FutureG AR/VR Applications
Lead: Robert LiKamWa
 

Topic: Scalable Render Streaming for Shared Immersive Environments
Presenter: Robert LiKamWa

Abstract: Immersive environments for education, training, and entertainment demand scalable VR/AR systems that support 20 to 200 headsets with high-fidelity, low-latency content. This talk benchmarks three key render streaming approaches: (1) independent game engine instances with pixel streaming, (2) shared game engines with multiple virtual cameras, and (3) point cloud multicast streaming with local rendering on headsets. We'll explore the performance of these scenarios, as well as advances in volumetric streaming formats, e.g., megaframes and multiresolution volumes of interest, to optimize bandwidth and computational resources. The discussion will highlight challenges and opportunities in foveated streaming, dynamic stream switching, and the role of 5G technologies in enabling efficient, large-scale immersive environments.


 

September 25, 2024

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Reducing communication cost of Split Federated Learning
Presenter: Chaitali Chakrabarti

Abstract: Split Federated Learning (SFL) is an emerging privacy-preserving decentralized learning scheme which splits a machine learning model between client and server such that most of the computations are offloaded to the server. Since the activations and gradients have to be communicated between the client and server, the communication cost of SFL is quite high. Earlier we had described a method to reduce the communication cost by compressing the transmitted data using bottleneck layers. In this talk we investigate the use of auxiliary networks, which allow for activations being sent periodically to the server, to reduce the communication cost. We show that a combination of bottleneck layers and auxiliary networks can achieve very high reduction in communication cost with minimal degradation in accuracy for both IID and non IID data. We also present our initial findings on using asynchronous model aggregation to reduce the training time in SFL for heterogeneous clients.

Topic 2: Improving performance of zone-specific CSI feedback method
Presenter: Chaitali Chakrabarti

Abstract: Zone-specific CSI feedback can be used to improve CSI encoding and decoding performance. Earlier we showed that for a given floor plan, the encoder model can be trained to be location-aware and can be shared across all zones. In this talk we describe our attempts at creating a zone floor plan that best ties in with the zone-specific training scheme. We improve the clustering-based floor plan generating method by including more features such as power-delay profile, spatial correlation, etc. and also tuning the feature weights. We show that the new floor plan helps further improve zone-specific CSI encoding and decoding performance.


 

September 18, 2024

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: Enhancing CSI-Based Location Authentication with Reconfigurable Intelligent Surfaces
Presenter: Yanchao Zhang

Abstract: Reconfigurable Intelligent Surfaces (RIS) have the potential to revolutionize future-generation wireless communication systems with applications such as improving signal strength and coverage, enabling adaptive beamforming, and boosting energy efficiency. In addition, RIS can be employed to design new physical-layer security mechanisms by manipulating the wireless environment in ways that are controllable by legitimate users but unpredictable to adversaries. In this presentation, we will share our ongoing work, which leverages RIS to enhance continuous location authentication (CLA) based on channel state information (CSI). CSI-CLA relies on the principle that each physical location has a unique multipath environment, resulting in verifiable, distinct CSI patterns. However, recent research presented at the 2024 IEEE Symposium on Security and Privacy-one of the top conferences in the field-demonstrates that adversaries can spoof these CSI patterns to defeat CSI-CLA systems. To address this challenge, we explore how RIS, in combination with advanced AI techniques, can strengthen CSI-CLA against such attacks. We will also present the promising results of our preliminary experiments using real RIS devices in our lab, which suggest the potential efficacy of this approach.


 

September 11, 2024

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Task Offloading Strategies in Resource-constrained Edge Tactical Networks
Presenter: Zhaofeng Zhang

Abstract: Recent years have witnessed explosive growth in deploying IoT devices over tactical edge networks. Due to the limited computing resources of local edge devices, there is an urgent need to offload the computing tasks to edge servers. A central question here is “How can we design offloading strategies and resource allocation plans so that all the computing resource constraints of servers and communication constraints between devices and servers are satisfied?” In this paper, we formulate the problem of jointly optimizing task offloading and resource allocation (JOA) to maximize the total task completion rewards subject to various resource constraints under two wireless communication schemes: orthogonal frequency-division multiple access (OFDMA) and time-division multiple access (TDMA). We designed two algorithms to solve the OFDMA-based JOA problem. The first is an exact algorithm using a search tree, where the branch-cutting criterion is well-crafted based on the property of the JOA problem. The second is a meta-heuristic algorithm based on the simulated annealing approach. Under the TDMA scheme, the JOA problem is a multiple multi-dimensional knapsack problem. We solve this problem using a dynamic programming (DP) approach with a pseudo-polynomial time complexity, assuming that the number of servers is bounded by a constant. We then decrease the time complexity of DP by leveraging the rounding and scaling techniques. We conduct numerical evaluations and demonstrate that the proposed exact algorithms can obtain the optimal task offloading strategy and resource allocation plan. The heuristic algorithms can efficiently provide competitive performance.


 

September 4, 2024

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Designing Near-Optimal Partially Observable Reinforcement Learning
Presenter: Ming Shi

Abstract: Partially observable Markov decision processes (POMDPs) have been widely applied in various real-world applications. However, existing results have shown that learning in POMDPs is intractable in the worst case. The main challenge lies in the lack of latent state information. For example, in wireless channel scheduling, due to energy and security constraints, it is usually difficult or impossible for the user to know the conditions/states of all channels. Thus, a key fundamental question here is: how much online state information (OSI) is sufficient to achieve tractability? In this talk, I will talk about the fundamental conditions and methods that we develop for bridging the gap between partially observable reinforcement learning and networking with incomplete state information. Specifically, (i) I will discuss a lower bound that reveals a surprising hardness result: unless we have full OSI, we need an exponentially scaling sample complexity to obtain an ε-optimal policy solution for POMDPs. (ii) motivated by the structures of practical systems, we identify important subclasses of POMDPs that are tractable, even with only partial OSI.


 

August 28, 2024

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Presenters: Antonia Papandreou-Suppappola & Yongho Seo

Abstract: We present our latest results on tracking multiple targets in complex environments. The tracking scene has an unknown and time-varying number of targets that can enter or leave the scene at any given time. A multimodal radar platform provides independent measurements from multiple returns from the targets; the association of each return to a target is assumed unknown. The tracking environment is complicated by the presence of unknown clutter and time-varying noise.

We first demonstrate the use of the dependent Dirichlet process to learn the unknown and time-dependent information on the target labels. The inherent dynamic dependencies in the target state transition follows from the fact that the label of a target at the current time step depends not only on the labels and number of targets that were previously considered at the same time step but also on the labels and number of targets on the previous time step. This is necessary in order to ensure that the previous and current state parameters correspond to the same target.

We consider two different environments. In the first environment, unknown clutter is embedded within the measurements and the noise characteristics are known. In this case, we use dependent DP mixture modeling to associate each measurement with a target and separate the measurements from the clutter. In the second environment, in addition to the unknown clutter, the Gaussian noise characteristics are unknown and time-varying. In this case, we compute the joint distribution of the association probabilities and the measurements using conditional marginal distributions and Gibb's sampling. We demonstrate our results using simulations.


 

August 14, 2024

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Presenter: Andrew Herschfelt

Abstract: To support the extreme data rates promised by high-frequency communications systems, phased antenna arrays have become a popular solution for efficiently directing energy towards a target receiver. Coupled with modern beamforming techniques, these arrays offer higher flexibility, smaller form factors, and lower production and maintenance costs compared to legacy solutions like dish antennas. As the supporting processors become more capable, modern digital beamforming techniques also offer better precision, control, and directionality than analog techniques, but consume more power due to the inclusion of additional analog-to-digital converters (ADCs). At high frequencies (10-100 GHz) the antenna elements shrink dramatically, so these arrays can support significantly more antenna elements (1,000-10,000). Digital beamforming techniques scale poorly in this regime; the power required to operate the ADCs becomes prohibitively expensive and data distribution across the elements becomes bottle-necked by Tb/s throughput requirements. To address these key limitations, we propose a novel hierarchical digital beamforming solution that significantly reduces the power consumption (10x) and throughput requirements (100x) with only minor performance degradation. We benchmark the proposed solution against ideal analog and digital beamforming solutions for reasonable array configurations in the Ku and Ka bands to quantify these reductions and demonstrate the feasibility of this technique in practical applications.


 

August 7, 2024

Thrust 7: FutureG AR/VR Applications
Lead: Robert LiKamWa
 

Topic: Enhancing Cross-Virtuality Geospatial Data Visualization For VR/AR and Screen-based Collaboration
Presenters: Lauren Gold and Robert LiKamWa

Abstract: Virtual reality (VR) and desktop interfaces offer unique perspectives for analyzing complex multi-scale, multi-layer geospatial data, yet they often struggle with communication barriers due to a lack of group awareness. Collaborators often explore the same data on vastly different scales, hindering efficient communication and group coordination tasks. This work explores cross-virtual interactions to facilitate a shared context for multi-scale collaborators to work together efficiently. We integrate traditional annotation techniques while also studying new cues such as scale-adaptive gaze sharing and VR-to-PC hand-waving. We present Common Ground, a networked overlay system that integrates with existing screen-based GIS software and immersive geospatial data environments, enabling users of both platforms to coordinate effectively. We conducted a pilot study around real-world geologic analysis scenarios with two different populations of experienced geospatial visualization practitioners. Our qualitative observations of user behavior and questionnaire feedback underscore Common Ground's efficacy in supporting cross-virtual group awareness for multi-scale collaborators.


 

July 31, 2024

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Split Federated Learning with heterogeneous clients
Presenter: Chaitali Chakrabarti

Abstract: Split Federated Learning (SFL) is an emerging privacy-preserving decentralized learning scheme which splits a machine learning model between client and server such that most of the computations are offloaded to the server. We consider a more realistic scenario where clients are heterogeneous, and the heterogeneity can be with respect to both their computing capability and communication capability. Earlier we had described a method that addresses communication-level heterogeneity through use of bottleneck layers (BL) and compute-level heterogeneity through processing different number of layers at the client end. We showed that our method had good accuracy performance for IID data. In this talk, we describe our attempts to handle training with non-IID data, specifically data that has label distribution skew. We focus on use of knowledge distillation techniques to help improve the training accuracy performance.

Topic 2: Improve zone-specific CSI feedback using multi-task learning
Presenter: Chaitali Chakrabarti

Abstract: Channel state information (CSI) encapsulates path characteristics information of a communication channel and is used to optimize transmission strategies in high data rate systems. Prior work proposed zone-specific CSI feedback to improve CSI encoding and decoding performance. Earlier we described a multi-task learning (MTL) method for the CSI feedback problem that integrated zone information into the training process. For a given floor plan, we showed that the encoder model can be trained to be location-aware and can be shared across all zones. In this talk we describe our attempts at creating a zone floor plan that best ties in with the zone-specific training scheme. We show the performance of a k-means clustering-based floor plan generating algorithm that uses location and power statistics as features. Next steps will include better clustering methods including adding more representative features.


 

July 24, 2024

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: SimSML: Simultaneous Spectrum Misuser Localization for Dynamic Spectrum Sharing
Presenter: Yanchao Zhang

Abstract: Spectrum misuser localization (SML) plays a crucial role in dynamic spectrum sharing to ensure that only authorized users can access and utilize the shared spectrum. In this presentation, I will introduce Simultaneous Spectrum Misuser Localization (SimSML), a novel framework that harnesses the power of Graph Neural Networks (GNNs) to achieve simultaneous, accurate, and rapid localization of multiple spectrum misusers, even when they are in close proximity. Our approach capitalizes on location-dependent RSS residues derived from the log-distance path loss model. Through our inductive GNN model, we predict real-time RSS residues for potential transmitter-receiver pairs. During system runtime, these predicted RSS residues, combined with real-time RSS measurements, enable efficient transmission-power estimation and transmitter localization using the zero-forcing method. Comparing SimSML with the latest SML schemes, we observe significant advantages, achieving 37~58% lower localization errors and 14~24% lower transmission-power estimation errors across various scenarios with comparable computational time and overhead. These results demonstrate the promising potential of SimSML as an advanced and effective SML solution.


 

July 10, 2024

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Throughput-guaranteed Resilient Routing Scheme: A Conditional Value at Risk Approach
Presenters: Zhaofeng Zhang and Guoliang Xue

Abstract: We will continue our discussion on throughput-guaranteed resilient routing schemes. Inspired by the recent traffic engineering work TeaVar, we take a step further to answer the following questions: 1) Can we achieve robust routing of mission-critical flows without pre-computed paths for each connection request? 2) Can we incorporate network resource cost constraints into this robust routing scheme? We take a conditional value-at-risk (CVaR) approach to this study. Specifically, we propose an optimization problem to minimize the CVaR of the throughput loss across all the failure cases while satisfying the constraints of network bandwidth capacities and resource costs. Unfortunately, directly solving the above problem will induce cyclic flows. This challenge does not exist in the traffic engineering approach because the routing paths are pre-computed. To tackle this issue, we propose a bisection approach to obtain an upper bound of the CVaR so that the corresponding routing strategy is acyclic. We further discuss the relationship between the routing approaches by minimizing the CVaR of throughput loss and maximizing the expected throughput. Preliminary numerical results provide insights into the relationships among various approaches and the advantages and limitations of each approach.


 

July 3, 2024

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Efficient Reinforcement Learning for Wireless Communications
Presenters: Ming Shi and Mohamed Abd-Elmagid

Abstract: We will present our recent progress on RL with human feedback (RLHF) and RL for reducing age of information (AOI). Specifically, we will first present RL with human feedback. Traditional RL usually assumes a perfect reward function a priori. However, in many practical scenarios, this is apparently not true. To resolve this issue, reinforcement learning with human feedback (RLHF) has been causing attention in wireless communications. Motivated by this, in this work, we consider the setting where multiple users provide feedback based on their personal reference reward model on-the-fly. Due to the inconsistency and/or unreliability of human feedback, there could be discrepancy between these reference models and the ground-truth reward model. Our preliminary results explain the value of human feedback in robust RL, even if they are corrupted and online. In addition, we will present how RL can be useful to provide provably efficient algorithms for networks with imperfect information. In this work, we develop a provably efficient episodic RL algorithm that optimizes this tradeoff between the transmission cost of a source node (with limited battery capacity) and the achievable AOI, in the case of having unknown channel statistics. When the action is to drop the generated status update in a certain time slot (which occurs when the AoI is relatively low), it would be useful to utilize that slot for exploration at a lower power cost (by sending dummy packets/pilot signals). The proposed algorithm surprisingly achieves an order-optimal regret (i.e., O(1)) with respect to the number of episodes.


 

June 26, 2024

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Topic: Agile waveform design for dynamic spectrum sharing
Presenter: Antonia Papandreou-Suppappola

Abstract: We will present the following work on agile waveform design for dynamic spectrum sharing:
- Waveform co-design for radar and wireless communications systems by optimizing multiple, and possibly conflicting, objective functions
- Waveform design for radar co-existence with wireless communications systems using different approaches: (1) Control theoretic approach based on waveform-dependent estimation error bounds; (2) Reinforcement learning approach based on time-varying environmental or operational conditions; (3) Time-frequency feature extraction and clustering approach.

We will continue our discussion on Bayesian tracking with nonparametric learning (following our April 10 2024 presentation). We will review the use of clustering using the dependent Dirichlet process to track multiple targets and provide new simulations on tracking under unknown and time-varying measurement noise conditions.


 

May 22, 2024

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Split Federated Learning with heterogeneous clients
Presenter: Chaitali Chakrabarti

Abstract: Split Federated Learning (SFL) is an emerging privacy-preserving decentralized learning scheme which splits a machine learning model between client and server such that most of the computations are offloaded to the server. While existing SFL schemes focus on homogeneous clients, a more realistic scenario is when clients are heterogeneous, and the heterogeneity can be with respect to both their computing capability and communication capability. Earlier we had described a method to address communication-level heterogeneity through use of bottleneck layers (BL) that support different degrees of data compression. Here we also consider compute-level heterogeneity where different clients process different number of layers in the model based on their compute capabilities. Aggregating heterogeneous client models with different number of layers and different sized BLs is quite challenging. We propose a method for heterogeneous model aggregation that has accuracy comparable to the case with communication heterogeneity with a small increase in computation cost at the client-end.

Topic 2: Improve zone-specific CSI feedback using multi-task learning
Presenter: Chaitali Chakrabarti

Abstract: Channel state information (CSI) encapsulates path characteristics information of a communication channel and is used to optimize transmission strategies in high data rate systems. Prior work proposed zone-specific CSI feedback to improve CSI encoding and decoding performance. By decomposing the CSI feedback task into multiple zone-specific subtasks, each of the zone-specific autoencoder models have lower complexity thereby reducing the computation cost at the user equipment (UE) end. However, such a system requires the UE to store multiple encoder models leading to a large memory footprint. We propose a multi-task learning-based (MTL) method for the CSI feedback problem that integrates zone information into the training process. By adding an extra index predictor task to the MTL framework, the encoder model is trained to be location-aware and can be shared across all zones. We show that the proposed method enables the CSI feedback to benefit from zone-specific training while achieving a low computation cost and low memory footprint at the UE end.


 

May 15, 2024

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic: Securing Contrastive mmWave-based Human Activity Recognition against Adversarial Label Flipping
Presenter: Yanchao Zhang

Abstract: Wireless Human Activity Recognition (HAR), leveraging their nonintrusive nature, has the potential to revolutionize various sectors, including healthcare, virtual reality, and surveillance. The advent of millimeter wave (mmWave) technology has significantly enhanced the capabilities of wireless HAR systems. This paper presents the first systematic study on the vulnerabilities of mmWave-based HAR to label flipping poisoning attacks in the context of supervised contrastive learning. We identify three label poisoning attacks on the contrastive mmWave-based HAR and propose corresponding countermeasures. The efficacy of the attacks and also our countermeasures are experimentally validated on a prototype system. The attacks and countermeasures can be easily extended to other wireless HAR systems, thereby promoting security considerations in system design and deployment. This paper has been accepted and will be presented in ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec), Seoul, Korea, May 27-30, 2024. If time permits, I will also briefly discuss our ongoing work on utilizing Reconfigurable Intelligent Surfaces (RIS) for various cybersecurity applications within FutureG systems.


 

April 24, 2024

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Presenter: Guoliang Xue

Abstract: This presentation will cover two topics.
(1) Resilient Network Resource Allocation: Achieving robust routing of mission-critical flows is important over tactical edge networks. We seek to answer a natural question: "Can we route mission-critical flows to simultaneously satisfy the flows' demands and survive when network failure happens?" We take a conditional value-at-risk (CVaR) approach to this problem. We propose a multi-objective optimization problem to concurrently minimize the upper bound of the gap between CVaR of the arbitrary flow's throughput and its demand while maximizing the total network throughput across all the failure cases. To establish the tractability of this problem, we demonstrate its convexity. To deal with the non-differentiability of the original formulation, we further transform a non-linear non-differentiable convex optimization problem into an equivalent linear programming problem, which can be solved within polynomial time. Our preliminary numerical results show that the proposed approach can satisfy the network demands as much as possible when network failure happens.
(2) Task Offloading Strategies in Tactical Networks: In this on-going research, we study task offloading strategies in wireless tactical networks. Each task has its corresponding deadline, CPU requirement, RAM requirement, and precision requirement. We need to decide whether to offload the task and where to offload the task. Since the offloading needs to go through wireless media, SINR is an important factor to ensure feasible and timely transmission. We formulate the optimization problem and derive some technical details.


 

April 17, 2024

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic: Design Machine Learning for Muti-Agent Systems: Fairness and Action Space Reduction
Presenters: Mohamed Abd-Elmagid and Peizhong Ju

Abstract: For multi-agent systems such as cross-layer network control, data-driven methodologies have the potential to significantly outperform traditional model-based approaches. Leveraging the power of machine learning, specifically reinforcement learning, our group has investigated some important topics taking into account the specific constraints and dynamics of wireless networks in tactical environments such as non-stationarity, safety and average soft constraints, partial observability of the environment, etc. In today's discussion, we will focus on two specific topics: achieving fairness across various users and alleviating the curse of dimensionality.

In the first part, we present a novel RL strategy to achieve fairness in multi-agent Markov Decision Processes (MDPs). This approach diverges from traditional value function maximization by introducing a fairness function, using online convex optimization within a confidence-bounded policy space to achieve sub-linear regret and provide a probably approximately correct (PAC) guarantee. We further propose an offline fair RL algorithm and provide theoretical performance guarantee.

The second part proposes a strategy to mitigate the curse of dimensionality in high-dimensional MDPs by intelligently grouping similar actions. This method balances performance degradation with efficiency by employing a linear decomposition model to analyze the trade-offs between grouping granularity and estimation errors.


 

April 10, 2024

Thrust 2: Dynamic and Agile FutureG Networks
Lead: Antonia Papandreou-Suppappola
 

Presenter: Antonia Papandreou-Suppappola

Abstract: The presentation will discuss two topics: (a) Bayesian tracking with nonparametric learning and (b) wireless system coexistence with waveform design and interference mitigation. The outline of the presentation is as follows.
* Overview of Dirichlet process and dependent Dirichlet process for clustering and mixture modeling.
* Tracking multiple targets under various time-varying scenarios, including unknown and time-varying target cardinality, unknown target labels, unknown measurement-to-target association and unknown time-varying measurement noise intensity.
* Waveform design for coexisting spectrum sharing systems.
* Interference mitigation in spectrum sharing environments.


 

April 3, 2024

Thrust 1: FutureG RF/mmWave/THz Networks
Lead: Dan Bliss
 

Presenters: Andrew Herschfelt and Dan Bliss

Abstract: Modern RF environments are becoming increasingly congested. This limits the opportunities and capabilities of modern RF systems, obstructing the development and proliferation of new technologies. Modern spectral convergence techniques promise enhanced system performance and a wider range of applications, but also require a fundamentally different design philosophy from traditional and legacy systems. Furthermore, these new technologies cannot ignore the significant infrastructure of existing technologies (telecommunications, radar, GPS, etc.) and must also be flexible enough to co-exist with them. We explore the performance limitations of novel and legacy multi-function systems in the context of co-existence and co-design. We derive performance bounds for several compelling applications and discuss how a transition to co-designed multi-function systems enables mutual performance enhancements and a wider range of sub-system capabilities.


 

March 27, 2024

Thrust 7: FutureG AR/VR Applications
Lead: Robert LiKamWa
 

Topic: Harnessing Volumetric Streaming for Athletic Performance Analysis
Presenter: Robert LiKamWa

Abstract: In the realm of remote sports coaching, traditional methods often fall short in providing comprehensive spatial analysis, a critical component for in-depth athletic performance evaluation. This talk delves into advancements in volumetric streaming technologies that are transforming the way coaches interact with and analyze athletes' performances from afar.

We introduce "Augmented Coach," a system we published/presented at IEEE VR '24, which leverages volumetric data to create detailed 3D representations of athletes in motion, as well as UX/UI tools for spatial annotation and visualization capabilities, facilitating a deeper understanding of movement patterns and enabling effective communication with athletes, regardless of their location. We will also explore the developments in streaming volumetric data transmission in real-time, particularly over 5G and FutureG networks. This technology is not only pivotal for sports analysis but also showcases the potential for broader applications in various fields requiring high-definition, real-time 3D data transmission.


 

March 13, 2024

Thrust 6: Energy-Efficient FutureG Heterogeneous SoC
Lead: Chaitali Chakrabarti
 

Topic 1: Split Federated Learning with heterogeneous clients

Abstract: Split Federated Learning (SFL) is an emerging privacy-preserving decentralized learning scheme suitable for resource-constrained devices. SFL splits a machine learning model between client and server such that most of the computations are offloaded to the server. Unfortunately, SFL has high communication cost since the activations and gradients have to pass back and forth between clients and server. To reduce the amount of activation/gradient data, we make use of data compression using bottleneck layers (BL). Furthermore, in a heterogeneous system, where the clients subscribe to different data rates, the training time of SFL can increase because of clients with low data rates. We address this problem by compressing data from low data rate clients more by using smaller BL compared to data from high data rate clients. We propose HetBL, a method that aggregates data from different sized BLs with minimal performance degradation. Results on image classification problems show that compared to competing methods, HetBL achieves significant training time reduction with the lowest accuracy loss for IID data and mild accuracy loss for non-IID data.

Topic 2: Network task classification with multi-task learning

Abstract: Network traffic classification identifies the type of application or activity of network traffic. Deep learning (DL) models have achieved state-of-the-art performance in traffic classification, but their development is constrained by the limited size of labeled network traffic data. Multi-task learning (MTL) helps enhance the performance of traffic classification by mitigating the large sized labeled data requirement through use of traffic information such as latency, duration, bandwidth, etc., that are easy to collect. However, traditional MTL based methods result in weak performance when the gradients of the individual tasks are not aligned. An evaluation of MTL using traffic classification and bandwidth prediction shows that up to 36% of the model updating steps suffer from gradient conflicts. To address this problem, we propose to use a gradient projection method which reduces the gradient conflicts and helps the combined gradient achieve faster model convergence. Preliminary results show that such a method has the potential to improve the performance of network traffic classification.


 

March 6, 2024

Thrust 5: Secure and Trustworthy FutureG Networks
Lead: Yanchao Zhang
 

Topic 1: Membership Inference Attack against RF Fingerprinting

Abstract: Deep learning-based RF fingerprinting (DRFF) systems have gained prominence for their effectiveness in wireless device authentication based on unique RF hardware features in wireless signals. However, the inherent vulnerabilities of deep learning (DL) models make DRFF systems susceptible to DL attacks tailored for RF fingerprinting. In this presentation, we will introduce W-MIA, the first practical label-only membership inference attack (MIA) against DRFF systems. W-MIA can passively eavesdrop on RF signals to construct a shadow model and perform MIA covertly. Additionally, it can enhance attack performance through active interaction with RF fingerprinting systems. We also propose a simple yet effective countermeasure against W-MIA. Extensive experiments confirm W-MIA's high attack efficacy in a label-only setting, comparable to the state-of-the-art which assumes a more knowledgeable adversary. Furthermore, our proposed defense matches the performance of existing defenses while minimizing usability loss in RF fingerprinting systems.

Topic 2: WaveKey: Secure Mobile Ad Hoc Access to RFID-Protected Systems

Abstract: We will also present the design and evaluation of WaveKey, a cross-modal deep learning-based method to enable mobile ad hoc in-situ access to RFID-protected cyber systems. Built upon the ever-growing popularity of user-carried mobile devices and RFID technologies, WaveKey is motivated by the need for secure and user-friendly data access in various application contexts. WaveKey explores a random gesture performed by the mobile user to induce correlated IMU data and RFID signals at the involved mobile device and RFID server, adopts deep learning techniques to extract the complex cross-modal correlation, and devises an Oblivious Transfer-based key-agreement protocol toward secure and efficient key establishment. Theoretical analysis and experimental human-based evaluations confirm the high security and efficiency of WaveKey.


 

February 21, 2024

Thrust 3: Scalable Network-Aware FutureG ML Control Mechanisms
Lead: Ness Shroff
 

Topic 1: Safety and Partial Observability in RL for Autonomous Networked Systems, as well as briefly discussion on some recent work on RL with human feedback
Topic 2: Online Learning for Optimizing Age of Information in Networks with Unknown Channels Statistics

Thrust 4: Intelligent and Resilient FutureG Network Management
Lead: Guoliang Xue
 

Topic: Critical Network Flow: A Conditional Value-at-Risk Approach

Abstract: Robust routing of mission critical flows is an important research problem in network science. Can we accommodate a new session without disturbing existing flows? If not, can we accommodate the new session with feasible mitigation of existing flows? How does the network respond in the event of a network element failure? We take a conditional value-at-risk (CVaR) approach to this problem. Our objective is to maximize the CVaR of network throughput under worst-case scenarios, particularly focusing on single link failures. Demonstrating the problem's tractability, we formulate it as a convex optimization problem. Nevertheless, numerous hurdles persist, primarily due to the non-differentiable nature of the problem.