Navigation

Program

The main conference will take place from November 23–25, 2026, in Tokyo, Japan. All times are in Japan Standard Time (JST). The program runs from 09:00 to 17:00 on Days 1 and 3 and from 09:00 to 16:55 on Day 2. A 105-minute midday break is scheduled on Days 1 and 2, and a 90-minute midday break is scheduled on Day 3.

The main conference follows a single-track format with ten technical sessions, each consisting of six papers. Each paper is allocated 13 minutes: 10 minutes for the presentation and 3 minutes for questions and discussion. Each 80-minute technical session concludes with a 2-minute wrap-up.

Click a paper title to view its abstract.

Monday, November 23, 2026

09:00 - 09:30 Opening Remarks
09:30 - 10:40 Keynote Speaker: Xiaojun Lin
Index Policies: Recent Advances, Future Challenges, and Opportunities for Managing an Agentic World
Xiaojun Lin

Title: Index Policies: Recent Advances, Future Challenges, and Opportunities for Managing an Agentic World

Speaker: Xiaojun Lin
Professor and Global STEM Scholar
Department of Information Engineering
The Chinese University of Hong Kong
Hong Kong SAR, China

Abstract: We are increasingly facing resource allocation problems involving many stateful agents, ranging from scheduling wireless transmissions to managing LLM inference tasks. When our performance objective is time-related, e.g., latency, freshness, or completion time, we often need to model the problem as a multi-agent MDP (Markov Decision Process), which unfortunately incurs exponentially high complexity when the problem size is large.

Index policies, e.g., the Whittle index, can be a useful tool to decompose an otherwise intractable multi-agent MDP into subproblems and produce low-complexity, scalable and adaptive solutions. However, the existing theory of Whittle index still has many severe limitations, preventing its use in many practically interesting scenarios.

In this talk, I will introduce some of our recent progress in expanding the use-cases for index policies, especially in dealing with situations where non-indexability arises and where there is a global source of randomness that affects all agents. I will conclude with some thoughts on how index policies could potentially be used to manage computation and communication of agentic workload, as well as on future challenges.

10:40 - 10:55 Coffee Break
10:55 - 12:15    Session 1: NextG Radio: Channels, Beams, and Network Measurements
Session Chair: TBD
OAM-Enabled Multi-User Uplink for Sub-THz Networks John Reko, Edward W. Knightly (Rice University) 10:55 - 11:08
Abstract: The sub-THz frequency band offers significant potential to increase wireless capacity, providing the capability to accommodate growing network demands. Multi-user uplink further improves spectral efficiency, enabling high-throughput communication from multiple users to a common access point. In this work, we propose MOUST , a sub-THz multi-user communication framework that leverages orbital-angular-momentum (OAM) multiplexing to enable high-rate communications. To adapt orthogonal broadside OAM systems for multi-user reception, we perform receive beam-steering for each user and embed each single-user phase profile into a multi-user receive aperture. We validate the proposed approach using two sub-THz experimental testbeds: one to characterize the amplitude and phase profiles across a wide bandwidth, and another to evaluate achievable data rates and multi-user gains. From these results we perform extensive analysis and observe an average 1.75x multi-user sum-rate gain compared to single-user transmission for two users. This provides insight into practical design trade-offs and quantifies the multi-user multiplexing capabilities of MOUST .
UACP: Uncertainty-Aware Channel Prediction in MIMO Networks Mohammad Abdi, Arman Elyasi (Northeastern University); Nathaniel Bastian (United States Military Academy); Francesca Meneghello, Francesco Restuccia (Northeastern University) 11:08 - 11:21
Abstract: The key issue with existing compression strategies for MIMO CSI feedback is that they rely on strong assumptions about channel statistics. Ultimately, this may lead to suboptimal CSI compression and, in turn, inaccurate MIMO precoding in case of model uncertainty and/or model misspecification. In this paper, we propose UACP, the first \textit{uncertainty-aware} CSI prediction and compression strategy for MIMO systems. UACP is rooted in imprecise probability theory and aims at dynamically selecting the lowest number of subcarriers for channel reporting to guarantee a low CSI prediction error at the transmitter for accurate precoding. UACP uses evidential regression to obtain an estimate of the CSI prediction error in addition to the prediction itself. The key intuition is that such an error is linked with epistemic and aleatoric uncertainty of the prediction process. These two components are effectively disentangled by UACP and used as a proxy to decide the compression level, i.e., the number of subcarriers to be fed back. The results show that UACP reduces feedback overhead by $92.7\%$ and improves throughput by $21.3\%$ relative to Wi-Fi ($N_g{=}4$), while incurring only a $0.33\%$ increase in EVM. UACP falls back to full feedback and even improves BER by $6.3\%$ under distribution shift.
Physics-Informed Kernelized Thompson Sampling for Beam Management in Millimeter-Wave Systems Sopan Sarkar, Marwan Krunz (University of Arizona) 11:21 - 11:34
Abstract: Millimeter-wave (mmWave) communication relies on highly directional beams, making links particularly sensitive to blockage and mobility. Maintaining connectivity therefore requires frequent beam probing and realignment, which incurs substantial overhead. We address this challenge by formulating beam tracking as a kernelized multi-armed bandit over the two-dimensional beam-angle domain and proposing GP-TSAF, a Gaussian-process Thompson sampling algorithm with a novel array-factor (AF) kernel. Derived directly from the known uniform planar array geometry, the AF kernel captures beam-pattern overlap across the codebook, including sidelobe interactions and null-induced decorrelation, without requiring learnable parameters. This structure allows each received signal strength (RSS) observation to update the posterior over the full codebook, enabling data-efficient tracking under limited measurements. To improve robustness under mobility, we further introduce adaptive correlation scaling, which reduces spatial smoothing when the observed RSS varies rapidly and enables faster adaptation without explicit mobility models. We establish cumulative and dynamic regret guarantees for GP-TSAF, with complexity governed by the effective dimensionality induced by the AF kernel rather than the codebook size. Experiments on both simulated and over-the-air datasets show that GP-TSAF consistently outperforms practical baselines in cumulative regret, achieved RSS, beam-identification accuracy, and recovery under abrupt channel changes.
Rethinking Passive RIS at Sub-6 GHz: A Reality Check on Practical Gains Fatemeh Salehi, Nagarjun Bhat (University of California San Diego); Manideep Dunna (UC San Diego); Dinesh Bharadia (UCSD) 11:34 - 11:47
Abstract: Reconfigurable intelligent surfaces (RISs) are widely viewed as low-power technologies for improving wireless links through programmable reflections. However, their practical utility at sub-6~GHz, where direct paths and rich multipath already exist, remains unclear. In this paper, we show that \emph{passive} RIS is not broadly beneficial in such environments. Through a combined simulation-and-hardware study of sub-6~GHz MIMO-OFDM systems, we find that its gains are highly conditional and often negligible under realistic illumination conditions. Our evaluation integrates Sionna ray-tracing simulations with over-the-air measurements using a ($12\times12$) dual-polarized passive RIS prototype consuming only $\sim$175~mW, tested in both indoor and outdoor settings under directional and low-gain transmitter configurations. Results reveal that passive RIS rarely improves the dominant channel mode. Instead, gains, up to 15~dB indoors and 12~dB outdoors, appear primarily in weaker secondary eigenmodes, indicating recovery of suppressed spatial degrees of freedom rather than enhancement of the main link. Under weak or omnidirectional illumination, the RIS often provides little benefit and can incur a 2--5~dB penalty. These findings demonstrate that passive RIS at sub-6~GHz is a highly conditional mechanism, whose effectiveness is governed by incident power density, antenna directivity, and channel geometry, rather than by the presence of the surface alone.
Measuring the Predictability of Beam-Level 5G Traffic:How Far Can Zero-Shot Foundation Models Go? Eduardo Baena, Phuc Dinh, Sizhe Wang, Dimitrios Koutsonikolas (Northeastern University) 11:47 - 12:00
Abstract: Beam-level traffic forecasting is critical for 5G resource allocation, yet the measurement properties of beam-level traffic remain poorly understood. We present a measurement-driven analysis across 2,961 time series from two public 5G datasets released by third parties. We first characterize how beam-level traffic differs from the aggregated cell-level traffic studied in prior work, identifying a sparsity-dependent predictability gradient: active beams exhibit strong periodicity and covariate correlations that enable accurate forecasting, while predictability attenuates monotonically with sparsity until all methods converge on the sparsest beams. We then benchmark 16 model configurations, including four general-purpose TSFM pre-trained on non-network data and one RAN-specific foundation model, across three experimental settings. A zero-shot TSFM significantly outperforms trained alternatives on 73\% of beams, but this advantage is concentrated on active beams where periodic structure is strongest; on the sparsest beams, foundation models hit the same noise-dominated predictability floor as ARIMA. The dominant factor is not model architecture, but input information: standard network telemetry covariates reduce error more than switching between any two models. Covariate-augmented models are also more resilient to distribution shift. Our oracle covariate setup represents an upper bound on this benefit. These results show that operators can replace per-beam trained pipelines with off-the-shelf foundation models augmented by readily available telemetry.
Experimental Analysis of Receiver Design for Near-Field Coordinated Beamforming Mars Akishev, Burak Bilgin, Edward W. Knightly (Rice University) 12:00 - 12:13
Abstract: Next-generation wireless networks are envisioned to operate in sub-THz bands to enable increased data rates and sensing resolution. In such bands, Coordinated Beamforming (CoBF) will be a critical technique to mitigate performance degradation that can occur when highly directional links are occluded by mobile environmental obstructions. Moreover, due to decreased wavelength, sub-THz networks can increasingly operate in the near field, which yields non-planar wavefronts at receivers. In this paper, we present the first experimental study of receiver (RX) beamforming for sub-THz CoBF operating in the near field. Specifically, we introduce a suite of RX beamforming policies that span Channel State Information at the Receiver (CSIR) requirements as well as RX beamforming capabilities. At one extreme, we define a baseline RX beamforming policy that uses an electromagnetic wave model to compute CSIR, accounting for transmitter location, beam type, and near-field propagation characteristics; the RX then compensates for the computed phase profile over each RX array element to maximize coherent combining. At the other extreme, we consider strategies that, despite operating in the near field, employ far-field beam steering via a linear phase gradient at the receiver. Such policies are not only simple in CSIR requirements (not requiring phase information for each antenna element) but also simplify receiver configuration, requiring only a single parameter to characterize a linear phase gradient. We first experimentally study an \emph{unobstructed} CoBF link and quantify the benefits of compensating for the nonlinear incident phase as compared to conventional far-field oriented methods. We then experimentally study \emph{obstructed} links and measure the performance of the policies under partial blockage, as well as RX power improvement from real-time adaptation to blockage.
12:13 - 12:15 Session Wrap-Up
12:15 - 14:00 Break
14:00 - 15:20    Session 2: Wireless Sensing, Localization, and Spatial Intelligence
Session Chair: TBD
BatStation: Toward In-Situ Radar Sensing on 5G Base Stations with Zero-Shot Template Generation Zhihui Gao (Duke university); Zhecun Liu, Tingjun Chen (Duke University) 14:00 - 14:13
Abstract: The coexistence between incumbent radar signals and commercial 5G signals demands a ubiquitous radar sensing for efficient and adaptive spectrum sharing. In this context, leveraging the densely deployed 5G base stations (BS) for radar sensing is particularly promising, offering both wide coverage and immediate feedback to 5G scheduling. However, the coexisting radar signals are superimposed on concurrent 5G uplink transmissions received by the BS, and practical deployment also demands a lightweight, portable radar sensing model. This paper presents BatStation, a lightweight, in-situ radar sensing framework seamlessly integrated into 5G BSs, which leverages uplink resource grids to extract radar signals through three key components: (i) radar signal separation to cancel concurrent 5G transmissions and reveal the radar signals, (ii) resource grid reshaping to align time–frequency resolution with radar pulse characteristics, and (iii) a zero-shot template correlation-based model that is trained purely on synthetic data that supports detection, classification, and localization of radar pulses, and can be directly applied to experiments. We implement BatStation on a software-defined radio (SDR) testbed and evaluate with real 5G traffic in the CBRS band. Results show robust performance across diverse radar types, achieving per-pulse detection probabilities of 97.02–79.23%, classification accuracy of 97.00%, and localization errors of 2.68–6.20 MHz (frequency) and 24.6–32.4 μs (time). Notably, BatStation requires runtime latency of only 0.11/0.94 ms on GPU/CPU, meeting the real-time requirement of 5G networks.
Continuous Passive Target Tracking over Intermittent and Narrow Sensing Beams in 6G Networks Fei Shang (The University of Hong Kong); Liang Liu, Jiannong Cao (The Hong Kong Polytechnic University); Chenshu Wu (The University of Hong Kong) 14:13 - 14:26
Abstract: 6G networks promise an advanced infrastructure for high-precision passive sensing and tracking, particularly in emerging sectors like Low-Altitude Economy (LAE) and Internet of Vehicles. However, realizing passive sensing in 6G remains challenging due to its unique signal characteristics, especially the extreme directionality of narrow beams and the intermittent availability of sensing resources, which are subordinated to the primary communication traffic. Even slight beam misalignments can lead to severe SNR degradation and consequently target loss. In this paper, we propose SPOT, a novel passive beam tracking framework that achieves continuous, high-precision tracking using only temporally sparse sensing slots. We employ an Interacting Multiple Model combined with an Extended Kalman Filter (IMM-EKF) to robustly predict non-linear target dynamics. We further develop a beam selection strategy that maximizes Fisher information gain to minimize tracking uncertainty, integrated with an opportunistic exploration mechanism to detect emerging targets. To reduce computation complexity, we introduce a lightweight approximation of the Posterior Cramér-Rao Lower Bound. Extensive evaluations using both simulations and real-world datasets demonstrate that SPOT significantly outperforms existing baselines. Our results show that SPOT achieves a detection rate near 100% and maintains a Root Mean Square Error (RMSE) of tracking below 1 meter, even in multi-target scenarios with highly sparse temporal updates.
Pushing Bistatic Wireless Sensing toward High Accuracy at the Sub-Wavelength Scale Wenwei Li, Jiarun Zhou, Qinxiao Quan (Peking University); Fusang Zhang (Beihang University); Daqing Zhang (Peking University) 14:26 - 14:39
Abstract: Contactless sensing using wireless communication signals has garnered significant attention due to its non-intrusive nature and ubiquitous infrastructure. Despite the promise, the inherent bistatic deployment of wireless communication introduces clock asynchronism, which leads to unknown phase offsets in channel response and hinders fine-grained sensing. State-of-the-art systems widely adopt the cross-antenna channel response ratio to cancel these detrimental phase offsets. However, the channel response ratio preserves sensing feature accuracy only at integer-wavelength target displacements, losing sub-wavelength fidelity. To overcome this limitation, we derive the first quantitative mapping between the distorted ratio feature and the ideal-channel sensing feature. Building on this foundation, we develop a robust framework that leverages raw channel response amplitude to recover the ideal sensing feature from the distorted ratio. Real-world experiments across Wi-Fi and LoRa signals demonstrate that our method can effectively reconstruct sub-wavelength displacement details, achieving nearly an order-of-magnitude improvement in accuracy.
Distributed Collaborative Positioning through ISAC for GPS-Denied Robotic Swarms Yu-Tai Lin, Ting-Wei Hsu, Karthik Sundaresan (Georgia Institute of Technology) 14:39 - 14:52
Abstract: Robotic swarms in GPS-denied environments require accurate position and velocity in real-time for all nodes, without external infrastructure or centralized control. Existing positioning approaches, including collaborative ones, fail at this due to non-scalable ($O(N^2)$) measurement overhead or centralized architectures that cannot keep pace with high-mobility swarms. We present \system, a novel paradigm of \emph{Di}stributed \emph{Co}llaborative\emph{P}ositioning through asynchronous, multi-static ISAC. Each node broadcasts a lightweight ISAC waveform, leveraging which, all neighbors passively and asynchronously measure relative distances and Doppler shifts through a novel differential sensing technique that cancels clock and channel frequency offsets with only $O(1)$ overhead. Coupled with a novel joint distance-velocity localization, each node independently estimates its position and velocity, alleviating topology rigidity requirements of conventional distance-based methods. An ISAC-aware distributed MAC further prioritizes fast-moving, under-measured nodes' measurements to accelerate convergence. Evaluated on our UGV swarm prototype, as well as NS3 simulations for large-scale UAV swarms (up to 100 UAVs at 15 m/s), DiCoP achieves sub-2 m error for over 89.38\% of estimates, 0.062 m/s velocity error, and 54.3 ms convergence time, outperforming GPS, IMU, WiFi FTM, and centralized baselines while maintaining stable accuracy over extended mission duration.
Simultaneous Infrastructure and Layout Mapping via Enterprise WiFi Yichen Wei, Yibo Pi, Yifei Sun (Shanghai Jiao Tong University); Yihui Li, Dingxiang Peng (Ruijie Networks); Cailian Chen (Shanghai Jiao Tong University) 14:52 - 15:05
Abstract: Environment-aware capabilities have been envisioned to enhance WiFi networks in communication, localization, and sensing. However, existing environment-aware approaches often assume prior knowledge of infrastructure (AP orientations and locations) and environment maps, limiting their applicability in real-world deployments. We present WISEMap, a framework that simultaneously reconstructs infrastructure and indoor environment maps via enterprise Wi-Fi, without prior information. The core challenge lies in the tight coupling between the two tasks: environment reconstruction demands precise infrastructure mapping, which in turn requires accurate mapping of the environment. WISEMap addresses this challenge by alternately optimizing the two tasks and progressively refining their estimates. Real-world evaluations show that WISEMap not only accomplishes both tasks simultaneously but also delivers substantially higher accuracy than the baselines designed for each task, reducing environment reconstruction error by 43–77% and the average AP and user localization errors by 70% and 39%, respectively.
Towards Agile and Active Scene Update for Embodied Space Exploration: A WiFi-Guided Approach Cheng Pan (Hong Kong University of Science and Technology); Cong Zhang (Simon Fraser University); Edith Ngai (The University of Hong Kong); Jiangchuan Liu (Simon Fraser University); Bo Li (Hong Kong University of Science and Technology) 15:05 - 15:18
Abstract: Timely scene updating is essential for embodied agents operating in dynamic indoor environments. While cameras and LiDAR are effective for high-fidelity reconstruction and verification, relying on them alone for long-term maintenance is costly, since robots must repeatedly revisit large areas to rediscover relatively sparse changes. This paper presents \textbf{WiDR}, a WiFi-guided framework for efficient and active scene updating. The key idea is to use ubiquitous wireless signals as a low-overhead prior indicating where meaningful changes are likely to occur, thereby guiding visual or depth-based verification, rather than using wireless sensing to reconstruct the scene itself. WiDR integrates wireless priors with floorplan context to build a belief-aware navigation graph, and formulates scene updating as a budget-constrained utility maximization problem that jointly considers navigation, sensing, updating, and model refinement. To remain effective in long-term dynamic environments, WiDR further incorporates active refinement using robot-verified observations to improve its change prediction quality. Experimental results in a three-room indoor testbed show that WiDR achieves higher update utility under the same budget by directing sensing effort toward high-value regions instead of relying on coverage-driven or vision-only exploration.
15:18 - 15:20 Session Wrap-Up
15:20 - 15:40 Coffee Break
15:40 - 17:00    Session 3: Scheduling, Latency, and Information Freshness
Session Chair: TBD
Optimal Parallel Scheduling under Concave Speedup Functions Chengzhang Li (The Ohio State University); Peizhong Ju (University of Kentucky); Atilla Eryilmaz (Ohio State University); Ness Shroff (The Ohio State University) 15:40 - 15:53
Abstract: Efficiently scheduling parallel computing resources across multiple concurrent jobs is a core challenge in modern cloud and edge computing systems that support AI-driven applications. Allocating more resources to a job accelerates its completion, but with diminishing returns. Prior work (heSRPT) has solved this problem for certain speedup functions with an exponential form, providing a closed-form solution. However, the general case with arbitrary concave speedup functions---which more accurately capture real-world workloads---has remained open. In this paper, we solve this open problem by developing optimal scheduling algorithms for parallel jobs under general concave speedup functions. We first discover a fundamental and broadly-applicable rule for optimal parallel scheduling, namely the \emph{Consistent Derivative Ratio (CDR) Rule}, which states that the ratio of the derivatives of the speedup functions across active jobs remains constant over time. To efficiently compute the optimal allocations that satisfy the CDR Rule, we propose the \emph{General Water-Filling (GWF)} method, a more general version of classical water-filling in wireless communications. Combining these insights, we design the \emph{SmartFill} Algorithm to solve the general scheduling problem. Unlike heSRPT, which always allocates resources to all active jobs, SmartFill selectively determines which jobs should receive resources and how much they should be allocated. For a broad class of so-called \emph{regular} speedup functions, SmartFill yields closed-form optimal solutions, while for non-regular functions it efficiently computes the optimum numerically. Simulation results show that SmartFill can outperform heSRPT across a wide range of concave speedup functions.
Asymptotically Optimal Joint Generation-Planning and Transmission-Scheduling for Periodic Deadline-Constrained Wireless Communication Jiaxin Huang, Xiaojun Lin (The Chinese University of Hong Kong); Guocong Quan, Jinbei Zhang (Sun Yat-sen University) 15:53 - 16:06
Abstract: We study the joint design of packet generation-planning and transmission-scheduling for periodic, deadline-constrained wireless communication. While transmission scheduling has been extensively studied in the literature, the generation planning problem has been largely neglected. However, a poor packet generation plan (e.g., when all sources generate new packets at the same time) causes high latency that cannot be overcome by transmission scheduling alone. We propose a new joint offline-online optimization framework for this problem, in which generation timing is optimized offline to spread arrivals evenly across time, while online transmission decisions are made via a Whittle index scheduling policy. Interestingly, we find that the Whittle index policy by itself (obtained through a standard relaxed problem) may not be optimal given an arbitrary generation plan. Instead, it becomes optimal only after we jointly design the generation plan (based on the solution to the same relaxed problem). We prove that our joint solution is asymptotically optimal as the system scales, by introducing new techniques that capture the periodic and time-inhomogeneous nature of the system and that overcome the difficulty due to the generation plans changing with the system scale. Simulations verify our optimality results and show that our joint solution achieves substantially lower latency plus deadline-violation costs than other baselines.
TSDM: A Scheduling Policy for Joint Throughput-AoI Optimization in Multichannel Wireless Networks Lin Wang, I-Hong Hou (Texas A&M University) 16:06 - 16:19
Abstract: Optimizing for both low Age of Information (AoI) and high throughput is critical for remote sensing applications that rely on multichannel wireless networks. However, jointly optimizing these two metrics is an analytically challenging problem, particularly in systems with heterogeneous and unreliable channels. To address this challenge, we propose TSDM, a Two-Stage Deficit Matching scheduling framework. TSDM is based on a second-order approach that characterizes the performance of each data flow by its mean and temporal variance. In the first stage, TSDM translates the high-level utility maximization objective into a concrete set of target mean and temporal variance statistics for transmissions over each node–channel pair. In the second stage, a low-complexity Weighted Matching Deficit (WMD) rule performs real-time channel assignment. We theoretically prove that TSDM achieves the desired mean and temporal variance for each flow. Furthermore, we conduct extensive simulations on two open joint throughput-AoI optimization problems. In both cases, TSDM significantly outperforms existing scheduling policies.
Windex: Index-Driven Contextual Scheduler for Scalable Service Guarantees in NextG Cellular Networks Archana Bura (Postdoctoral Scholar, UCSD); Ushasi Ghosh (University of California San Diego); Alexander Riedel (Texas A&M University); Dinesh Bharadia (UCSD); Srinivas Shakkottai (Texas A&M University) 16:19 - 16:32
Abstract: Next-generation cellular networks must support heterogeneous service objectives, ranging from throughput-oriented eMBB to latency- and regularity-constrained URLLC and XR traffic. Enforcing these objectives at the RAN scheduler is difficult because wireless channels, queues, and application arrivals evolve at the millisecond timescale while spectrum resources are shared across users. We formulate service-aware RAN scheduling as a constrained Markov decision process with per-user throughput and service-regularity constraints. A Lagrangian relaxation gives an exact decomposition of the relaxed problem into per-user control problems with a common resource price. We prove fixed-dual indexability for an illustrative TSLS-debt arm with a throughput objective and TSLS constraint, note the analogous scalar throughput-debt result, and assume fixed-dual indexability for the full vector state used in deployment. We design \emph{Windex}, an index-driven scheduler that learns dual-dependent per-user break-even resource prices using compact neural approximators over queue, channel, TSLS, and violation-deficit state. We further design \emph{CWindex}, a contextual extension that infers service class from queue-history time series when explicit labels are unavailable. We implement Windex as a real-time RIC $\mu$App on an open 5G testbed. Inference takes 10--20~$\mu$s per user on a general-purpose CPU, well within a 1~ms control budget. Trace-driven emulation using over-the-air channel measurements and larger simulations show that Windex and CWindex substantially reduce per-user service violations while maintaining throughput efficiency.
AOTP: Towards Application-Agnostic Network Transport for Data Freshness Zechao Chen (Zhejiang University); Jiayu Pan (Northeastern University); Tie Qiu (Tianjin University) 16:32 - 16:45
Abstract: Minimizing theAge of Information (AoI) is critical for time-sensitive applications, yet a gap still exists between AoI optimization theory and practical implementation. Existing solutions are often inherently constrained: application-layer approaches are tailored to specific use cases and lack generality, whereas link-layer solutions require hardware-specific changes that hinder deployment. To address these limitations, we introduce the Age-Oriented Transport Protocol (AOTP), a new transport-layer solution that offers application-agnostic AoI optimization. AOTP enforces a strict separation of concerns, hiding complex optimization logic from applications while requiring no kernel or hardware modifications. At its core, AOTP enhances the Robbins-Monro algorithm to efficiently adapt to dynamic network conditions in its Sampling Mode, and provides a Decoupled Mode that leverages delay-gradient pacing for general applications that generate samples at their own pace. Extensive ns-3 simulations across wireless links and multi-hop topologies demonstrate that AOTP consistently outperforms state-of-the-art baselines. Specifically, AOTP reduces average AoI by up to 99.4% under heavy multi-flow contention, mitigates bufferbloat, and preserves fairness for legacy cross-traffic under load.
Balancing Synchronization and Sum Delay in Cloud Virtual Reality Applications Shukai Chen, Koushik Kar, Alhussein A. Abouzeid (Department of Electrical, Computer, and Systems Engineering, Rensselaer Polytechnic Institute) 16:45 - 16:58
Abstract: This paper addresses the joint optimization of synchronization delay and total motion-to-photon latency in multiuser interactive cloud-based virtual reality systems. We propose a rendering strategy that switches each user between local and cloud rendering to balance synchronization and sum delay. While prior studies have optimized either inter-user synchronization delay or sum delay independently, we define and analyze the Balanced Delay Optimization (BDO) problem under a constraint on the number of cloud-rendered users. We prove that BDO is NP-hard in general. Under a cloud-faster-than-local condition, which induces a submodular regime, we identify a polynomial-time solvable unconstrained case and, for the NP-hard constrained case, develop a minimum $s$-$t$ cut-based approximation algorithm with a provable suboptimality bound, and show that it runs within a 90 Hz VR frame budget over the tested range. Numerical simulations and Unity-based prototype experiments validate the proposed framework. In particular, the proposed algorithm closely tracks the exact optimum in the submodular regime and remains competitive in a non-submodular violation-case simulation.
16:58 - 17:00 Session Wrap-Up
18:30 - 20:30 Reception

Tuesday, November 24, 2026

09:00 - 10:20    Session 4: Network Intelligence and Optimization
Session Chair: TBD
Routing and Topology Design for Low Earth Orbit Satellite Networks under Demand Uncertainty Chirag Rao (Massachusetts Institute of Technology); Eytan Modiano (MIT) 09:00 - 09:13
Abstract: We study the space segment of Low Earth Orbit (LEO) satellite networks that support satellite-to-satellite communication via inter-satellite links (ISLs). Our goal is to jointly design topologies and routing strategies that are robust to uncertain traffic demand. Leveraging the regularity and symmetry inherent in LEO constellations, we analyze traffic-oblivious routing schemes that exploit this structure. We show that Valiant Load Balancing (VLB) is optimal for symmetric networks in minimizing worst-case congestion. On the topology side, we demonstrate that minimizing average shortest path length (ASPL) leads to asymptotically optimal performance as the network scales. We introduce a new class of symmetric and uniform topologies, called 2-color topologies, which outperform conventional grid-based designs in terms of worst-case congestion and ASPL. In particular, our 2-color topologies achieve an ASPL scaling of $\Theta(\sqrt[3]{N})$, whereas traditional grid-based topologies have an ASPL of $\Theta(\sqrt{N})$ for a network of $N$ nodes. We validate these performance gains on a realistic-scale LEO constellation with practical link range constraints.
MEO: Mining Reliable Expert Signals for Offline Reinforcement Learning in Wireless Networks Lipeng Zu, Hansong Zhou, Yu Qian, Shayok Chakraborty (Florida State University); Yukun Yuan (University of Tennessee at Chattanooga); Linke Guo (Clemson University); Xiaonan Zhang (Florida State University) 09:13 - 09:26
Abstract: The operation of next-generation wireless networks increasingly hinges on Deep Reinforcement Learning (DRL) to optimize critical network decisions. Offline RL trains agents using accessible datasets and thus eliminates costly online interactions, rendering it particularly advantageous in dynamic wireless environments that demand low latency and high safety. However, its capability remains underexplored in wireless settings, where offline data are often limited and subject to partial observability. As a result, the learned policy fails to fully capture the informative decision patterns hidden in the dataset. To bridge this gap, we propose MEO, a framework that mines reliable expert signals from wireless datasets for offline RL deployment. MEO first identifies reliable expert samples from imperfect offline data, and then learns a structured latent representation to better distinguish expert and non-expert behaviors. Based on the representation, MEO further derives a compensable reward signal that helps offline RL algorithms more effectively exploit valuable expert knowledge. We evaluate MEO in three practical wireless network scenarios using datasets collected from real-world deployments. The results show that MEO consistently outperforms existing methods across all three scenarios. Notably, MEO can be seamlessly integrated with any offline RL algorithms, highlighting its generalizability and potential for wide adoption in wireless networks.
Foundation Model-Aided Multi-Agent Reinforcement Learning for Wireless Random Access Network Optimization Myeung Suk Oh (Georgia Southern University); Zhiyao Zhang (The Ohio State University); Alvaro Velasquez (University of Colorado Boulder); Nathaniel Bastian (United States Military Academy); Jia (Kevin) Liu (The Ohio State University) 09:26 - 09:39
Abstract: Random access (RA) is one of the most foundational medium access control (MAC) layer scheduling schemes for handling unpredictable data traffic from multiple terminals. While multi-agent reinforcement learning (MARL) has been explored to optimize RA-based wireless networks, its reliance on experience-driven, distributed policy learning incurs significant training overhead for each optimization task, limiting their feasibility in real-world applications. In this work, we propose to leverage a foundation model (FM) to improve MARL efficiency across diverse RA network optimization tasks. Specifically, we design an FM-aided actor-critic algorithm within a consensus-based decentralized MARL architecture and provide its convergence analysis under local reward exchanges and nonlinear value function approximations to show that our algorithm achieves the same convergence order as the conventional MARL with critic model exchanges and linear approximations. Our numerical results show that our FM-based approach significantly enhances MARL speed for RA network optimization.
Model-based Zero-Shot Conflict Detection Framework via Gradients in O-RAN Zeyu Fang, Shu Hong (George Washington University); Huu Trung Thieu, Nakjung Choi (Nokia Bell Labs); Tian Lan (George Washington University) 09:39 - 09:52
Abstract: Open Radio Access Network (O-RAN) enables network control through multi-vendor xApps operating both within and across layers, subnets, and domains, whose concurrent execution can trigger conflicts that are latent during the development phase. Existing conflict management approaches rely heavily on joint-execution data, which is often unavailable in practice. To address this limitation, we formalize a novel problem termed conflict reasoning, which involves identifying conflict-inducing conditions given only marginal datasets from each individual xApp. We propose ZODIAC, a three-stage framework for zero-shot conflict condition inference that comprises uncertainty-aware surrogate model training, trajectory-level diffusion training, and compositional guided denoising for efficient, physics-constrained, and reliable condition search. We derive a theoretical lower confidence bound showing that the compositional reasoning in ZODIAC serves as a principled surrogate for true conflict severity, with the epistemic penalty directly controlling the approximation gap. We evaluate ZODIAC on both the lightweight Mobile-Env platform across all three O-RAN Alliance conflict types (direct, indirect, and implicit) and a realistic NS-O-RAN-Flexric simulator. ZODIAC consistently outperforms baseline condition search methods, achieving over 20% higher True Positive Rate at Top-20, substantially stronger Spearman rank correlation, greater scenario diversity, and competitive computational efficiency. Ablation studies confirm the necessity of each guidance component, with epistemic uncertainty penalties proving essential for filtering spurious conflicts. To the best of our knowledge, ZODIAC is the first framework in O-RAN that enables conflict reasoning from marginal offline data without requiring any joint-execution traces.
TWINERGY: Digital Twin Empowered Practical Energy Saving in Heterogeneous Mobile Networks Yibo Ma, Tianxin Wang (The University of Edinburgh); Haoye Chai (Beijing University of Posts and Telecommunications); Yong Li (Tsinghua University); Mahesh Marina (The University of Edinburgh) 09:52 - 10:05
Abstract: Energy consumption has become a pressing concern as mobile networks expand rapidly, with the radio access network accounting for the dominant share. Among energy-saving techniques, cell shutdown stands out for its large saving potential: it selectively switches off capacity cells in heterogeneous networks while a coverage layer preserves service. However, optimizing cell shutdown policies faces two key challenges: (i) live trial-and-error is prohibitively risky, while traditional simulators cannot capture the complex dynamics needed for reliable evaluation; (ii) operator-deployed threshold-based policies are interpretable but extremely hard to optimize at scale. On the other hand, binary optimization is more tractable but yields policies incompatible with operator workflows. We propose TWINERGY, a digital twin-empowered framework for practical cell shutdown at scale. TWINERGY builds a data-driven Network Digital Twin for safe virtual evaluation, and decouples policy design into two stages: first optimizing binary on/off decisions under QoS constraints, and then mapping them into deployable threshold configurations, with a dedicated module for each. For scalable decision optimization, we design CABLE, a multi-agent reinforcement learning framework with shared policies, causal graph-based spatio-temporal encoding, and credit-guided cooperative learning. For threshold synthesis, we design a differentiable mapping module that introduces SoftControl as a surrogate for hard dual-threshold logic, enabling thresholds to be calibrated in continuous space via stable supervised optimization. Experiments on a digital twin built from 14,349 real-world base stations and deployments across three commercial trial networks with over 3,400 cells show that TWINERGY achieves up to 26.0% and 12.4% energy savings in 4G and 5G digital-twin evaluation, respectively, and delivers up to 13.07% and 12.20% energy-efficiency improvements in rural and dense 5G live networks without degrading QoS.
Morpheus: Interactive RF Digital Twins for In-Situ Layout Orchestration in Dynamic mmWave Environments Arko Datta (Indian Institute of Technology Madras); Akash Kumar Paul, Ayon Chakraborty (Indian Institute of Technology, Madras) 10:05 - 10:18
Abstract: Millimeter-wave (mmWave) networks provide high wireless capacity, but remain fragile in indoor environments, where small changes in clutter or object placement disrupt dominant propagation paths and cause cascading degradation. Automated RF planners optimize throughput but ignore operational constraints or intents that are difficult to enumerate or encode \emph{a priori}, often producing infeasible configurations. We present Morpheus, an interactive, human-in-the-loop framework for proactive, in-situ layout orchestration in mmWave environments. Automated approaches optimize RF performance but ignore intent, while human operators enforce feasibility but lack RF awareness. Morpheus bridges this gap through a human-in-the-loop design, using an RF-aware digital twin to identify plausible future configurations and recommend layout adjustments that improve throughput while preserving intent through learned priors and user feedback. We evaluate Morpheus on two real-world industrial deployments. Morpheus improves throughput by 25--40\% over human-driven layouts while reducing deviation from feasible configurations by 2--3$\times$, and achieves comparable throughput to automated planners with significantly fewer intent violations.
10:18 - 10:20 Session Wrap-Up
10:20 - 10:40 Coffee Break
10:40 - 12:00    Session 5: Distributed Learning and Collective Communication
Session Chair: TBD
A Tale of Two Learning Algorithms: Multiple Stream Random Walk and Asynchronous Gossip Peyman Gholami, Hulya Seferoglu (University of Illinois Chicago) 10:40 - 10:53
Abstract: Although gossip and random walk-based learning algorithms are widely known for decentralized learning, there has been limited theoretical and experimental analysis to understand their relative performance for different graph topologies and data heterogeneity. We first design and analyze a random walk-based learning algorithm with multiple streams (walks), which we name asynchronous “Multi-Walk”. We provide a convergence analysis for Multi-Walk w.r.t number of iterations (computation), wall-clock time, and communication cost. We also present a convergence analysis for “Asynchronous Gossip”, addressing the notable gap in the existing literature by explicitly accounting for its computational and communication overhead. Our results show that Multi-Walk has better convergence in terms of iterations as compared to Asynchronous Gossip in graphs with large diameters (e.g., cycles), while its relative performance, as compared to Asynchronous Gossip, depends on the number of walks and the data heterogeneity in graphs with small diameters (e.g., complete graphs). In wall-clock time analysis, we observe a linear speed-up with the number of walks and nodes in Multi-Walk and Asynchronous Gossip, respectively. Finally, we show that Multi-Walk outperforms Asynchronous Gossip in communication overhead, except in small-diameter topologies with extreme data heterogeneity. These results highlight the effectiveness of each algorithm across varying graph topologies and levels of data heterogeneity. Our codes are available for reproducibility.
Optimizing Stochastic Gradient Push under Broadcast Communications Tuan Nguyen, Ting He (Pennsylvania State University) 10:53 - 11:06
Abstract: We consider the problem of minimizing the convergence time for decentralized federated learning (DFL) in wireless networks under broadcast communications, with focus on mixing matrix design. The mixing matrix is a critical hyperparameter for DFL that simultaneously controls the convergence rate across iterations and the communication demand per iteration, both strongly influencing the convergence time. Although the problem has been studied previously, existing solutions are mostly designed for decentralized parallel stochastic gradient descent (D-PSGD), which requires the mixing matrix to be symmetric and doubly stochastic. These constraints confine the activated communication graph to undirected (i.e., bidirected) graphs, which limits design flexibility. In contrast, we consider mixing matrix design for stochastic gradient push (SGP), which allows asymmetric mixing matrices and hence directed communication graphs. By analyzing how the convergence rate of SGP depends on the mixing matrices, we extract an objective function that explicitly depends on graph-theoretic parameters of the activated communication graph, based on which we develop an efficient design algorithm with performance guarantees. Our evaluations based on real data show that the proposed solution can notably reduce the convergence time compared to the state of the art without compromising the quality of the trained model.
Over-the-Air Federated Learning in Heterogeneous Mobile Wireless Networks Ming Xiang (Northeastern University); Nicolo Michelusi (Arizona State University); Yonina Eldar (Northeastern University; Weizmann Institute of Science); Lili Su (Northeastern University) 11:06 - 11:19
Abstract: Over-the-air computation has emerged as a scalable and efficient solution for deploying federated learning algorithms in wireless networks by exploiting waveform superposition for simultaneous model aggregation. Most existing work struggles with heterogeneous fading channels. These approaches either enforce unbiased updates from all devices or allow partial device contributions while requiring careful tuning of the convergence bound to mitigate bias under specific fading models. However, the former significantly amplifies receiver noise due to the weakest channel, whereas the latter is sensitive to fading model mismatch and converges only to a biased objective. Such bias can substantially degrade federated learning performance if not handled properly. To tackle these challenges, we propose FedOAG, which employs algorithmic components to automatically satisfy energy constraints via gradient normalization and evenly mix devices' updates through implicit gossiping. Importantly, FedOAG does not require transmission from all devices, nor does it rely on a specific fading model or knowledge of time-varying statistical channel distributions. We show that FedOAG converges to a stationary point of an unbiased non-convex objective at the best possible rate ${\mathcal{O}}(1/\sqrt{T})$ for any stochastic first-order method. We corroborate our analysis with numerical experiments over dynamic wireless conditions on real-world datasets.
Joint Optimizing Federated Learning and Unlearning over Wireless Networks Dongyu Wei, Hanzhi Yu (University of Miami); Yuchen Liu (North Carolina State University); Christo Kurisummoottil Thomas (Worcester Polytechnic Institute); Shiwen Mao (Auburn University); Mingzhe Chen (University of Miami) 11:19 - 11:32
Abstract: In this paper, a communication and energy-efficient federated learning (FL) and unlearning (FU) co-designed framework is proposed. In the considered model, multiple edge devices collaboratively train a model. Users are allowed to quit the FL training and remove their data impacts on the trained model. Hence, an FU training is performed after the FL training. To reduce the energy and time consumption of FU, the users are grouped into multiple clusters such that the users in a group will retrain a deep learning model when the users in the group request to quit the FL training. The base station (BS) needs to determine the number of clusters, the users in each cluster, and the resource block (RB) allocated to each user, according to the probability that each user requests to quit the FL, data distribution, wireless channel conditions, and time and energy constraints. This problem is formulated as an optimization problem whose goal is to minimize the expected loss by determining the number of clusters, user cluster assignment, and RB allocations. This problem is challenging to solve since 1) the BS cannot predict which users will request data deletion in advance and 2) no explicit equation exists to capture the relationship between the number of clusters, user cluster assignment, and the expected FU loss. To solve the formulated problem, we first analyze the convergence rate of FU to build an explicit relationship among the FU training loss, the number of clusters, and the users in each cluster. Based on this relationship, the original problem is simplified to a mixed-integer optimization problem which can be divided into the user cluster assignment and the RB allocation optimization subproblems. These two problems are alternately solved by the designed Dinkelbach's method and the fast iterative shrinkage-thresholding algorithm (Dinkelbach-FISTA) and the standard Hungarian algorithm. Simulation results show that our method improves classification accuracy by up to $11.1\%$ and achieves up to $3\times$ faster convergence compared to gradient-based and ISTA-based methods.
From Prediction to Decision: Dynamic Maintenance Scheduling in Industrial Edge Computing via Federated DRL Jialin Tian (Department of Computer Science, University of Exeter); Harry Min (Department of Computer Science and Technology, University of Cambridge); Jia Hu (Department of Computer Science, University of Exeter) 11:32 - 11:45
Abstract: In industrial edge computing, numerous heterogeneous devices operate continuously under varying environmental conditions. Their degradation not only leads to localized failures but also reduces overall system uptime, making proactive maintenance necessary. Existing solutions typically predict each device's remaining useful life using deep learning models and then rely on predefined thresholds or heuristic rules to trigger maintenance. However, this open-loop paradigm cannot dynamically adapt its decisions to evolving conditions and struggles to scale maintenance intelligence across distributed edge sites, especially under data privacy constraints. This paper transitions from prediction to decision by reformulating maintenance scheduling as a sequential decision-making problem. We design a long-term operational cost minimization (LTOCM) objective based on a novel model of the device lifecycle and maintenance actions. We further formulate LTOCM as a Markov decision process, introducing a degradation-aware reward on top of the cost-based reward to guide the agent toward cost-effective maintenance policies. Based on this formulation, we propose FDDM, a federated DRL scheme for dynamic maintenance in industrial edge computing. Edge-resident agents independently interact with local devices, collect maintenance trajectories and update policies through entropy-regularized proximal optimization with generalized advantage estimation. To coordinate cross-site policy learning without exposing raw data, agents periodically synchronize with the cloud-level orchestrator in a federated manner, using a momentum-based global update to stabilize convergence. Extensive experiments on the NASA C-MAPSS dataset with varying numbers of distributed agents demonstrate that FDDM achieves near-optimal maintenance decisions while improving training efficiency over baseline methods.
Bridge: Optimizing Collective Communication Schedules in Reconfigurable Networks with Reusable Subrings Anton Juerss (Weizenbaum Institut & Technische Universität Berlin); Stefan Schmid (Technische Universität Berlin & Weizenbaum Institute) 11:45 - 11:58
Abstract: Optical circuit-switched networks have emerged as an appealing alternative to electrical fabrics as they can reconfigure the network topology at runtime, reducing communication cost and improving bandwidth utilization. Yet exploiting optical reconfigurable networks for collective communication comes with a fundamental trade-off: each reconfiguration incurs non-negligible delay, communication must pause while the fabric reconfigures, and the benefit of a new topology depends on future traffic. The central question is therefore when reconfiguration is worth its cost. While prior work has demonstrated the benefits of reconfiguration, existing strategies use optical links only to optimize the current step, without \textit{reusing} them for future steps. In this paper, we present BRIDGE, a reconfiguration strategy for important collective communication primitives used in AI/ML and HPC applications, namely All-to-All, AllReduce, Reduce-Scatter, and AllGather. BRIDGE exploits the structure of Bruck's communication pattern to support efficient sparse reconfiguration. The key idea is to reduce propagation and transmission delay by directly connecting immediate communication partners and preserve efficient reachability to future peers through connected \textit{subrings}. As a result, optical links can be \textit{reused} across multiple subsequent steps, allowing the benefit of reconfiguration to \textit{amortize} beyond a single step. Our evaluation shows that BRIDGE reduces All-to-All completion time by typically 3x to 10x over static baselines even with millisecond-scale reconfiguration delays. For AllReduce, BRIDGE uniformly outperforms existing reconfiguration strategies, delivers up to 1.5x speedup, and exceeds the bandwidth-optimal ring algorithm by 1.5x to 6.6x on low to moderate-sized workloads.
11:58 - 12:00 Session Wrap-Up
12:00 - 13:45 Break
13:45 - 15:05    Session 6: Edge LLM Serving and Fine-Tuning
Session Chair: TBD
Serving Chain-structured Jobs with Large Memory Footprints with Application to Large Foundation Model Serving Tingyang Sun, Ting He (Pennsylvania State University); I-Hong Hou (Texas A&M University) 13:45 - 13:58
Abstract: As a current trend in Artificial Intelligence (AI), large foundation models are increasingly employed as the core of AI services. However, even after training, serving such models at scale remains a challenging task due to their heavy resource footprints, particularly in terms of GPU memory. While recent works revealed unique characteristics of systems serving foundation models that distinguish them from traditional distributed computing systems, there is still a lack of fundamental understanding of the underlying system management problems. This work aims at addressing this gap by extracting a novel problem of "server chain composition" via block placement and cache allocation for serving chain-structured jobs with large memory footprints, which models a fundamental problem in serving large foundation models through pipeline parallelism. After showing the NP-hardness of the optimal solution, the focus is turned to developing scalable algorithms with guaranteed performance under state-of-the-art load balancing. Application of the proposed solution to a distributed large language model (LLM) serving system shows significant reduction of response times compared to state-of-the-art solutions.
POLAR: Online Learning for LoRA Adapter Caching and Routing in Edge LLM Serving Shaoang Li, Jian Li (Stony Brook University) 13:58 - 14:11
Abstract: Edge deployment of large language models (LLMs) increasingly relies on libraries of lightweight LoRA adapters, yet GPU/DRAM can keep only a small resident subset at a time. Serving a request through a non-resident adapter requires paging its weights from storage, incurring measurable latency. This creates a two-timescale online control problem: on a slow timescale, the system selects which adapters remain resident in fast memory, while on a fast timescale it routes each request to an adapter whose context-dependent utility is unknown a priori. The two decisions are tightly coupled: the cache determines the cost of exploration, and the router determines which adapters receive informative feedback. We formulate this joint caching-and-routing problem as a two-timescale contextual bandit and propose POLAR (Paging and Online Learning for Adapter Routing). POLAR pairs a cache-aware LinUCB router with an epoch-based cache controller. We study two variants. A fixed-epoch version provides a robust baseline with worst-case regret guarantees under arbitrary contexts. An epoch-doubling version, POLAR+, adds forced exploration and improved cache optimization to achieve $\widetilde{\mathcal{O}}(d\sqrt{NT}+\sqrt{KT})$ sublinear regret under stochastic regularity and cacheability conditions, where $N$ is the adapter count, $K$ the cache size, $d$ the context dimension, and $T$ the horizon. The routing term matches the standard contextual-bandit rate up to logarithmic factors, showing that the memory hierarchy does not fundamentally slow routing learning. Experiments using 15 real LoRA adapters for Qwen2.5-7B together with measured GPU paging latencies show that adaptive cache control substantially outperforms non-adaptive baselines and exhibits scaling trends consistent with the theory.
Continuous Semantic Caching for Low-Cost LLM Serving Baran Atalar (Carnegie Mellon University); Xutong Liu (University of Washington); Jinhang Zuo (City University of Hong Kong); Siwei Wang, Wei Chen (Microsoft Research Asia); Carlee Joe-Wong (Carnegie Mellon University) 14:11 - 14:24
Abstract: As Large Language Models (LLMs) become increasingly popular, caching responses so that they can be reused by users with semantically similar queries has become a vital strategy for reducing inference costs and latency. Existing caching frameworks have proposed to decide which query responses to cache by assuming a finite, known universe of discrete queries and learning their serving costs and arrival probabilities. As LLMs' pool of users and queries expands, however, such an assumption becomes increasingly untenable: real-world LLM queries reside in an infinite, continuous embedding space. In this paper, we establish the first rigorous theoretical framework for semantic LLM response caching in continuous query space under uncertainty. To bridge the gap between discrete optimization and continuous representation spaces, we introduce dynamic $\epsilon$-net discretization coupled with Kernel Ridge Regression. This design enables the system to formally quantify estimation uncertainty and generalize partial feedback on LLM query costs across continuous semantic query neighborhoods. We develop both offline learning and online adaptive algorithms optimized to reduce switching costs incurred by changing the cached responses. We prove that our online algorithm achieves a sublinear regret bound against an optimal continuous oracle, which reduces to existing bounds for discrete query models. Extensive empirical evaluations demonstrate that our framework approximates the continuous optimal cache well while also reducing computational and switching overhead compared to existing methods.
MOSAIC: Efficient Collaborative Inference for Multi-Modal Large Language Models with Distributed Edge Computing Kongyange Zhao (Sun Yat-sen University; China Tobacco Guangdong Industrial Co., Ltd.); Zhixuan Liao (Sun Yat-sen University); Shengyuan Ye (Guangdong Power Grid Co., Ltd.); Tao Ouyang (Central South University); Liekang Zeng (The Chinese University of Hong Kong); Xu Chen (Sun Yat-sen University) 14:24 - 14:37
Abstract: Multi-modal large language models (MLLMs) enable natural language reasoning over heterogeneous sensory data, appealing to many edge AI scenarios, yet their deployment faces competing demands on latency, energy, and quality. We present MOSAIC, which adopts a component-level decomposition and distributed deployment paradigm for efficient MLLM inference at the edge, realizing three core ideas: parallel encoding, token pruning, and device-server collaboration. Specifically, for the offline stage, modality-specific encoders distilled from a unified foundation model are co-deployed with Q-Former projection on edge devices for parallel encoding, while LoRA instruction fine-tuning adapts the frozen LLM backbone to downstream tasks. For the online stage, we formulate a multi-objective optimization over latency and energy that jointly tunes token budgets, device frequencies, and quantization precision under long-term semantic accuracy constraints, achieving dynamic token pruning and device-server collaboration. We leverage Lagrangian relaxation to decompose the original problem into tractable per-slot subproblems and design an online algorithm with provable near-optimal performance and guaranteed constraint satisfaction. We implement a realistic testbed for extensive performance evaluation, which demonstrates that MOSAIC reduces latency by up to $49.8$\% and energy by up to $65.1$\% over advanced baselines while maintaining high semantic accuracy.
Just Talk Once: Communication-Efficient Split Federated LLM Fine-Tuning on Edge Devices Jiaxiang Geng (Duke Kunshan University); Xianhao Chen (The University of Hong Kong); Bing Luo (Duke Kunshan University) 14:37 - 14:50
Abstract: Large language model (LLM) fine-tuning is increasingly shifting toward data generated on edge devices, where memory, computation, bandwidth, and connectivity constraints make conventional federated learning difficult to sustain. Split federated fine-tuning (SFT) improves client-side efficiency by offloading most model parameters and computation to the server but requires step-by-step bidirectional communication loop across the split interface and forces continuous client involvement throughout training. In this paper, we present \textbf{L-shaped SFT}, a split fine-tuning framework that removes this bidirectional bottleneck. Our key insight is that weight tying in modern LLMs enables server-side hidden activations to be directly supervised using target embeddings, allowing the training loss to be computed on the server without returning server outputs to the client. To further eliminate the need for continuous client participation, based on L-shaped SFT, we introduce \textbf{one-shot SFT}, in which clients upload activations once and then go offline while the server continues optimization over cached representations. We implement our design in a real system testbed with heterogeneous edge clients, including commercial smartphones and NVIDIA developer boards. Experiments demonstrate that our schemes significantly reduce communication costs and client online time compared with existing SFT baselines.
FrozenSFL: Model-Frozen Split Federated Learning for Resource-Efficient LLM Fine-Tuning over Heterogeneous Edge Devices Xiaopei Chen (South China University of Technology); Wen Wu, Liang Li (Peng Cheng Laboratory); Fei Ji (South China University of Technology); Xu Chen (Sun Yat-sen University); Xuemin (Sherman) Shen (University of Waterloo) 14:50 - 15:03
Abstract: Collaborative fine-tuning of large language models (LLMs) at the network edge enables privacy-preserving and personalized generative intelligence, but is constrained by the limited memory and compute resources of edge devices. Traditional approaches, such as split federated learning (SFL), partition models between devices and a server, requiring iterative exchanges of activations/gradients while the server manages multiple submodels. This leads to substantial communication overhead, high activation storage costs, and severe server-side memory pressure when scaling to massive heterogeneous devices. In this paper, we propose \textbf{FrozenSFL}, a resource-efficient SFL framework for collaborative fine-tuning of LLMs across heterogeneous edge devices. At its core, FrozenSFL strictly freezes device-side submodels to eliminate both device-side backpropagation and server-to-device gradient transmission, thereby removing on-device activation storage requirements and downlink communication overhead. Building upon this design, we improve the framework in two aspects. \textit{First}, we design a layer-wise LoRA aggregation strategy to handle structural heterogeneity introduced by device-specific model partitioning and enable efficient server-side updates. \textit{Second}, we introduce a server-side feature caching mechanism that reuses invariant features, allowing subsequent training epochs to proceed offline and reducing redundant computation and communication. We implement the FrozenSFL system as an asynchronous multi-threaded pipeline, which serves as our real-world testbed for evaluation. On the device side, forward computation and data transmission are decoupled via a producer–consumer mechanism. On the server side, a first-in-first-out scheduling strategy serializes fine-tuning tasks across devices. A hybrid memory–disk streaming mechanism further buffers incoming data and ensures stable training under network variability. Extensive evaluations with representative models, including BERT-Base (0.11B), ViT-Large (0.3B), and LLaMA-2-7B (7B), validate the effectiveness of FrozenSFL. In particular, for LLaMA-2-7B, FrozenSFL can reduce device/server VRAM footprint by up to 14\%/82.7\% with only 1\% accuracy loss, while achieving up to 6.7× fine-tuning speedup.
15:03 - 15:05 Session Wrap-Up
15:05 - 15:35 Coffee Break
15:35 - 16:55    Session 7: Communication-Efficient Edge Intelligence
Session Chair: TBD
Semantic Multiplexing Mohammad Abdi, Francesca Meneghello, Francesco Restuccia (Northeastern University) 15:35 - 15:48
Abstract: Mobile devices require the parallel execution of several processing tasks at the edge to reduce their computing load. However, existing communication systems only support parallel transmissions at the bit level. In this paper, we introduce the new concept of \textit{Semantic Multiplexing}, which shifts stream multiplexing from \textit{bits} to \textit{tasks}. By merging multiple task inputs into a single \textit{semantic representation}, Semantic Multiplexing can multiplex more tasks than the number of physical streams by adding a (semantic) degree of freedom to the communication system. Semantic Multiplexing has been prototyped in an experimental testbed with a Jetson Orin Nano and millimeter-wave software-defined radios. Both image classification (vision) and sentiment analysis tasks (language) have been considered to evaluate Semantic Multiplexing against state-of-the-art baselines. Our experiments indicate that Semantic Multiplexing reduces latency, energy consumption, and communication load by up to 8$\times$, 25$\times$, and 54$\times$, respectively, compared to existing baselines while maintaining comparable performance. For example, image classification accuracy drops only by less than 4\% when increasing from 2 to 8 the number of tasks multiplexed over a 4$\times$4 channel. We pledge to publicly share the complete software codebase and the collected datasets for reproducibility. A video demonstration of Semantic Multiplexing is also available\footnote{\url{https://youtu.be/P25gDeOxtU0}}.
Communication-Aware Distributed Edge Inference via Latent Representation Compression Peyman Gholami (University of Illinois Chicago); Theodoros-Thirimachos Davarakis (Northeastern University); Teng Li (University of Illinois Chicago); Miquel Sirera (Northeastern University); Salil Reddy (Ohio State University); Ayberk Yarkın Yıldız (Northeastern University); Anish Arora (The Ohio State University); Atilla Eryilmaz (Ohio State University); Stratis Ioannidis (Northeastern University); Chengzhang Li (The Ohio State University); Hulya Seferoglu (University of Illinois Chicago); Ness Shroff (The Ohio State University) 15:48 - 16:01
Abstract: We study the problem of optimizing the performance of distributed model inference under constrained edge resources. We propose a framework that optimizes the trade-off between model accuracy and communication costs by controlling latent representation compression to meet strict Quality of Service (QoS) throughput targets. For settings with known channel state information (CSI), we derive a unique closed-form optimal solution for single tasks and reduce the multi-task problem to a convex optimization program characterized by a per-link water-filling strategy. We extend these to handle unpredictable environments via an Estimated Stochastic Dual Descent algorithm that relies only on causal channel estimates. We provide rigorous Lyapunov-based proofs demonstrating that our approach strictly satisfies long-term delay constraints while achieving a bounded optimality gap. Our results offer a robust, scalable blueprint for maximizing the performance of pipelined AI tasks in dynamic, resource-constrained distributed systems. We verify the effectiveness of our proposed framework through simulations and experiments with real edge devices.
Queue-Stable Online Learning: Bilinear Bandits for Semantic Codecs in NTNs Shijing Yuan (China Telecom Research Institute); Mengru Wang, Jie Li (Shanghai Jiao Tong University); Peng Li (Xi'an Jiaotong University); Hongyang Chen (Research Center for Intelligent Network); Zhi Liu (The University of Electro-Communications); Qun Jin (Waseda University); Song Guo (Hong Kong University of Science and Technology) 16:01 - 16:14
Abstract: Learned speech codecs can carry voice over narrowband non-terrestrial links, but profile selection couples the quality of the current speech segment to the sender backlog faced by later segments. Profile quality changes with the channel and representation parameters, while first transmissions and retransmissions share the modem grant. We present the Lyapunov Bilinear Upper Confidence Bound scheduler, LyBUCB, which combines a bilinear contextual bandit with Lyapunov control. The bandit learns selected-profile quality from online feedback; the controller combines its confidence score with the current payload and virtual distortion queues. A transport ledger records fresh payload, retransmissions, in-flight payload, and unique acknowledgments separately. We bound finite-horizon decision error relative to a drift-plus-penalty oracle evaluated on the same predecision state. For independent and identically distributed innovations under strict feasibility, exact current-state search yields the queue and online-score bounds; delayed ranking adds an explicit approximation term. Measurements on a Tiantong prototype and closed-loop tests with recorded traces show that LyBUCB gives higher session speech quality than five baselines and meets the preset queue and distortion limits. The measured observation-to-encoder and scheduler P95 latencies, together with the maximum score-return interval, are below the 500-ms slot duration.
PRISM: Reinforced Semantic Communication for Decentralized Federated Anomaly Detection in Vehicular Networks Georgios Levis (Unaffiliated); Antonio Di Maio (ETH Zurich) 16:14 - 16:27
Abstract: Decentralized Federated Learning (DFL) enables collaborative learning that preserves privacy, but faces critical communication challenges in bandwidth-constrained environments such as vehicular networks. We propose Policy-driven Representation and Information Selection Method (PRISM), a Reinforcement Learning (RL)-based method that optimizes semantic communication in DFL through adaptive representation selection and intelligent topology management. Our dual-model architecture separates task learning (autoencoder-based anomaly detection) from communication control (Proximal Policy Optimization (PPO)-based decision making), enabling agents to dynamically select between transmitting full model weights or compressed trajectory embeddings. We formulate this as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and train using a Centralized Training Decentralized Execution (CTDE) paradigm. We demonstrate our method through a digital twin monitoring scenario where vehicles collaboratively detect trajectory anomalies in neighbors' behavior, applicable to forensic analysis, insurance claims, and fleet auditing. Experiments on Simulation of Urban Mobility (SUMO) generated Barcelona traffic data show an F1-score of 0.90, scaling from 2 to 128 agents with graceful degradation past $N = 32$, Byzantine-robust aggregation, and validation on NGSIM US-101 real traces (F1 0.82). The learned policy exhibits emergent temporal behavior: transmitting weights early during training when model divergence is high, then shifting to embeddings as models converge, with the weight-to-embedding mix derived as the equilibrium of the communication budget constraint. The method generalizes to other bandwidth-limited collaborative learning domains.
CREWS: Collaborative Robust Edge WiFi Sensing with Asynchronous and Incomplete Observations Yinan Chen, Yang Zhou, Xiaoxia Huang (School of Electronics and Communication Engineering, Shenzhen Campus of Sun Yat-Sen University, Shenzhen, China); Pan Li (Hangzhou Dianzi University) 16:27 - 16:40
Abstract: Existing collaborative WiFi sensing systems rely on perfect node synchronization and complete data availability. However, real-world edge deployments suffer from heterogeneous computing and network dropouts, leading to asynchronous and incomplete features. We propose CREWS, a robust collaborative sensing framework that inherently resists these network volatility. First, CREWS employs a topology-agnostic aggregator invariant to the arrival order and subset size of incoming features. Second, rather than discarding delayed observations, it utilizes a staleness-aware adaptive replay mechanism. By treating stale features from lagging nodes as system-induced hard samples, CREWS transforms synchronization delays into beneficial training regularization. We theoretically prove the joint convergence of this architecture and demonstrate how replay bounds the bias-variance trade-off. Extensive evaluations and an 8-node heterogeneous hardware testbed demonstrate its superior resilience. Under severe conditions i.e., 50\% transient dropout rate or out-of-distribution jitter, CREWS restricts accuracy degradation to merely 2.2 percentage points, substantially outperforming state-of-the-art baselines.
SHEAR: Goal-Oriented Anomaly Detection on Compressed Time Series via Semantic Transition Graph Guoyou Sun (Aarhus university); Panagiotis Karras (University of Copenhagen); Qi Zhang (Aarhus University) 16:40 - 16:53
Abstract: Sensor data in IoT systems are commonly compressed to meet resource constraints, yet anomaly detection directly on compressed data remains largely unexplored. This challenge is particularly evident for recurrent anomalies, which commonly appear as repeated abnormal patterns in long-running IoT streams. Existing approaches often become cumbersome or expensive since they decouple compression from analytics, requiring detection to re-process full raw data, while supervised deep learning alternatives further demand labels or incur computational overhead unsuitable for edge deployment. In this paper, we propose SHEAR, a goal-oriented unsupervised method via Semantic Transition Graph (STG) induced directly from compressed time series. The STG organizes compressed representation as semantic states and encodes their temporal transitions, so that recurrent anomaly patterns are identified through unlikely state transition chains without decompression or neural network inference. Experiments on public datasets show that SHEAR, achieves better detection performance than state-of-the-art approaches, while running 3.9X to 11.2X faster than subsequence-oriented methods and up to 2219X faster than deep learning baselines, with substantially reduced storage and memory footprint.
16:53 - 16:55 Session Wrap-Up
18:30 - 20:30 Banquet

Wednesday, November 25, 2026

09:00 - 10:20    Session 8: Online Learning, Resource Allocation, and Network Economics
Session Chair: TBD
Restless Bandits with Individual Penalty Constraints: Near-Optimal Indices and Deep Reinforcement Learning Nida Zamir, I-Hong Hou (Texas A&M University) 09:00 - 09:13
Abstract: This paper investigates the Restless Multi-Armed Bandit (RMAB) framework under individual penalty constraints to address resource allocation challenges in dynamic wireless networked environments. Unlike conventional RMAB models, our model allows each user (arm) to have distinct and stringent performance constraints, such as energy limits, activation limits, or age of information minimums, enabling the capture of diverse objectives including fairness and efficiency. To find the optimal resource allocation policy, we propose a new Penalty-Optimal Whittle (POW) index policy. The POW index of an user only depends on the user's transition kernel and penalty constraints, and remains invariable to system-wide features such as the number of users present and the amount of resource available. This makes it computationally tractable to calculate the POW indices offline without any need for online adaptation. Moreover, we theoretically prove that the POW index policy is asymptotically optimal while satisfying all individual penalty constraints. We also introduce a deep reinforcement learning algorithm to efficiently learn the POW index on the fly. Simulation results across various applications and system configurations further demonstrate that the POW index policy not only has near-optimal performance but also significantly outperforms other existing policies.
Learning When to Update: A Near-Optimal Timing Bandit Approach Qiulin Lin, Junyan Su, Liyuan Wang (City University of Hong Kong); Minghua Chen (The Chinese University of Hong Kong (Shenzhen) and City University of Hong Kong) 09:13 - 09:26
Abstract: Systems operating in dynamic environments require timely updates to sustain performance. For resource-intensive systems such as machine learning models and digital twins, strategically timing updates is essential. Updating too frequently wastes resources, while updating too infrequently leads to costly performance degradation. The problem is particularly challenging when the system's degradation pattern is unknown a priori, as is common in new operating environments. We formalize this challenge as a novel \emph{timing bandit} problem, where each arm represents a candidate update interval with a fixed update cost and an unknown, stochastic degradation cost. Three structural properties distinguish this setting from standard multi-armed bandits: selecting an interval commits the learner to multiple time slots before the next update; arm costs are composed of per-step degradation costs and a fixed update cost; and selecting a longer interval naturally reveals degradation at every intermediate step, providing consecutive feedback relevant to shorter intervals. By exploiting these structures, we develop Balanced Consecutive Arm Elimination (BCAE). BCAE achieves $\tilde{O}(\sqrt{T})$ regret, improving upon the $\tilde{\Omega}(K\sqrt{T})$ regret of standard bandit algorithms in this setting, where $K$ is the number of candidate update intervals. We further propose an Optimism-Enhanced variant (OE-BCAE) that integrates lower-confidence-bound principles to improve empirical adaptivity while preserving the same regret order. Moreover, the regret bound achieved by our algorithms matches the theoretical lower bound up to logarithmic factors. Simulation results demonstrate that our algorithms achieve low regret and remain stable as both the number of arms and the update cost vary.
Reusable Capacity Management under Adversarial Arrivals: Learning with Delayed Feedback Qingsong Liu, Mahsa Sahebdel, Ativ Joshi, Mohammad Hajiesmaili (University of Massachusetts Amherst) 09:26 - 09:39
Abstract: We study an online reusable resource allocation problem motivated by mobile edge AI platforms, where sequentially arriving requests must be immediately admitted and assigned to one of multiple edge service options with limited reusable capacities or offloaded to a remote central server. If accepted, a request occupies one service slot of the selected option for a random duration and yields a stochastic reward. The problem is particularly challenging because arrivals may be adversarial, reward and service-time statistics are both unknown and heterogeneous across request types and service options, and reward feedback can be delayed beyond service completion. To address these challenges, we propose an index-based online learning algorithm that combines optimistic reward estimates with a conservative, utilization-aware opportunity-cost penalty that captures the future value of reusable capacity. Using a primal–dual framework built on a novel reverse-order benchmark, we establish competitive-ratio guarantees under both bounded and light-tailed unbounded feedback delays. In particular, under a homogeneous reward-service model with i.i.d. inter-arrival times, our algorithm achieves an asymptotic competitive ratio of $1-1/e$ as reusable capacities grow large. Experimental results on a real-world dataset further demonstrate the effectiveness of our approach.
Online Joint Pricing and Scheduling in Distributed Networked Platforms Mahsa Sahebdel, Ali Zeynali, Adam Lechowicz (University of Massachusetts Amherst); Bo Sun (University of Ottawa); Mohammad Hajiesmaili (UMass Amherst) 09:39 - 09:52
Abstract: In distributed networked platforms, such as geographically distributed data centers and electric vehicle charging networks, there are multiple resource providers that can serve user requests arriving sequentially over time. In this paper, we study an online posted pricing problem jointly with scheduling in distributed networked platforms. In such systems, user requests arrive sequentially over time, and each request (if accepted) can be served by one of multiple resource providers. Rather than centrally assigning a provider, the platform posts prices together with corresponding scheduling decisions for all providers, and each user independently selects one option or leaves based on the posted prices and its private valuation. As demand increases, posted pricing and scheduling act as complementary mechanisms for regulating system usage. However, designing effective online posted pricing policies is challenging because of uncertainty in arrivals and the intertemporal coupling across requests, where current pricing decisions affect future feasibility and performance. We propose a class of parameterized threshold-based posted pricing policies that enable tractable online decision-making with provable performance guarantees. We identify a sufficient condition that ensures feasibility and supports competitive analysis, which in turn characterizes a structured policy design space. Building on this structure, we develop a learning-based approach to improve empirical performance under realistic demand patterns. We evaluate our approach on the problem of pricing and scheduling in electric vehicle charging stations using real-world charging traces, and demonstrate that the empirical performance of our algorithm outperforms other benchmark algorithms.
Joint Incentive Contract and Assignment Design for Mobile Networked Platforms with Endogenous Supply Yumou Liu, Zhenzhe Zheng (Shanghai Jiao Tong University); Jiang Rong, Ling Yan (Xiaohongshu); Fan Wu, Guihai Chen (Shanghai Jiao Tong University) 09:52 - 10:05
Abstract: Mobile networked platforms such as crowdsensing markets, edge-computing exchanges, vehicular service systems, and sponsored-content programs must coordinate two coupled levers: an assignment rule that decides who serves each request and an incentive contract that determines how realized value is shared. In all four settings, supply is endogenous because providers remain active only when expected earnings cover their private operating cost. We study a discounted platform problem in which assignment and revenue sharing jointly shape future supply. We calibrate the tractable one-period objective using an auxiliary memoryless discounted surrogate: it maximizes a global lower bound on that surrogate objective, and near high-survival stationary operating points its retention coefficient becomes the local shadow price of discounted future welfare within the surrogate model. This yields an Equimarginal Principle requiring efficient platforms to equalize the marginal retention efficiency of assignment and contracting. Under complete information, we obtain an urgency-adjusted assignment score, a KKT characterization of the optimal contract, and a Frank-Wolfe coordinate-ascent benchmark with an $O(1/K)$ joint stationarity rate. Under hidden provider quality, we develop SE-RW-UCB, a subsidized-exploration algorithm with retention-weighted contextual assignment and a learned safe contract floor. A KuaiRec-based online recommendation benchmark shows that separated designs leave a sizable retention gap. In that benchmark, Oracle Fixed is the profit leader, whereas SE-RW-UCB attains the strongest high-type and all-provider retention at an interior contract. The service-quality differences are comparatively small, indicating that the main leverage of co-design in this calibration lies in supply preservation rather than per-request quality gains.
Decision-Focused Learning for Bidding Optimization in Online Advertising Qilong Lin, Zhenzhe Zheng (Shanghai Jiao Tong University); Zixiao Wang, Yan Jin, Tianxu Ji, Wei Wang (Tencent Inc.); Fan Wu, Guihai Chen (Shanghai Jiao Tong University) 10:05 - 10:18
Abstract: In large-scale networked systems such as online marketplaces and advertising exchanges, real-time decision-making under dynamic environments is a fundamental challenge. A widely adopted paradigm is the predict-then-optimize workflow, where the prediction stage estimates key factors for decisions, and the optimization stage optimizes objectives based on these factors. However, existing prediction models are typically trained for learning factors, without awareness of the downstream optimization objective, leading to suboptimal results. We investigate this problem in the context of first-price auction bidding, a canonical decision problem in networked platforms. Specifically, we first propose Decision-Focused Bidding (DFB) approach, which retains the predict-then-optimize workflow, but integrates the downstream objective into the training of the prediction model. To overcome the zero-gradient issue in back-propagation, we develop a smoothed decision mapping that provides usable gradients while preserving clear interpretability in the bidding context. We further show that DFB's decision-learning objective and the existing objective of learning winning probabilities are complementary under non-i.i.d. environments, and accordingly propose Enhanced DFB (EDFB), which jointly optimizes both objectives to achieve robust generalization in real-world bidding. Finally, our EDFB approach has been deployed in a production auto-bidding system, where a 3-day live A/B test over ~100 billion requests achieves a 2.3\% GMV lift online and a 16\% surplus improvement offline.
10:18 - 10:20 Session Wrap-Up
10:20 - 10:40 Coffee Break
10:40 - 12:00    Session 9: Security, Privacy, and Trust in Networked Systems
Session Chair: TBD
mmPoison: Targeted Clean-label Poisoning Attacks on mmWave-based Human Activity Recognition Haoran Cao, Jiadi Yu, Yanmin Zhu (Shanghai Jiao Tong University); Hong-Ning Dai (Hong Kong Baptist University); Linghe Kong (Shanghai Jiao Tong University, China) 10:40 - 10:53
Abstract: Nowadays, mmWave-based human activity recognition (HAR) systems are widely used in smart homes, healthcare monitoring, and telemedicine. To achieve high accuracy, HAR systems require large-scale data collection from distributed sources for model training, such as edge devices or public repositories, which opens up opportunities for attackers to inject poisoned data into the training process, inducing false predictions and damaging the reliability of mmWave-based HAR systems. This paper presents targeted clean-label poisoning attack strategies, mmPoison, which injects imperceptible poisoned data into the training datasets for causing misclassifications on target instances in mmWave-based HAR systems without controlling the labeling process. In mmPoison, different attack methodologies are proposed for different attacker capabilities. For white-box models, mmPoison updates perturbations based on parameter trajectory alignment, and then optimizes the poisoned instances to remain close to the clean instances. This process is iteratively applied to generate poisoned datasets. For black-box models, mmPoison trains surrogate models and then uses them in a generator–discriminator framework to generate poisoned data by jointly optimizing attack effectiveness and imperceptibility. Experiments conducted on public and self-collected datasets validate the effectiveness and imperceptibility of mmPoison in causing misclassifications within mmWave-based HAR systems
Your Palm Is in My Hand: Palm-Vein Biometric Spoofing via Electromagnetic Side Channel Ruoqin Tang, Xiusheng Cai, Yongliang Chen (City University of Hong Kong); Tao Ni (King Abdullah University of Science and Technology); Guowen Xu (University of Electronic Science and Technology of China); Qingchuan Zhao (City University of Hong Kong); Cong Wang (City University Hong Kong) 10:53 - 11:06
Abstract: As a biometric modality, palm-vein authentication has seen growing adoption in recent years and is becoming prevalent in wallet-free payments and critical access-control applications. Although palm-vein patterns are widely regarded as nearly impossible to forge or steal, our study reveals that the electromagnetic (EM) emanations induced during sensor operation can leak information correlated with a user's palm-vein characteristics and attackers can use it to reconstruct palm-vein images and spoof authentication systems. In this paper, we present PVSpy, a novel attack framework that demonstrates the feasibility of this attack. Specifically, PVSpy passively captures EM emanations radiated from palm-vein sensors, extracts features that map the captured signals to palm-vein images, and exploits diffusion models for biometric reconstructions to spoof palm-vein authentication devices. We extensively evaluate PVSpy on five commercial devices, and the recovered vein patterns exhibit a promising level of similarity. In addition, we conduct 30 end-to-end spoofing attacks against these commercial-off-the-shelf (COTS) devices with different types of palm-vein sensors, and PVSpy achieves 24.3% (top-1) and 63.3% (top-3) success rates.
Sharing of Irreversible Latent Features in lieu of RF IQ Data for Multiple Downstream Learning Tasks AADHITHYA; AADHARSH, Nasim Soltani (The University of Texas at Austin); Joydeep Ghosh (UT Austin); Kaushik Chowdhury (University of Texas Austin) 11:06 - 11:19
Abstract: Wireless in-phase and quadrature (I/Q) data is useful for training various radio frequency (RF) machine learning (ML) models. However, sharing such I/Q data faces barriers from concerns of (i) potentially violating laws related to recording raw samples that can eventually be decoded by an unauthorized third party, and (ii) revealing information about a proprietary system or the environment involved in the data capture. This paper proposes a different solution to sharing I/Q data intended for use in ML models by leveraging latent embeddings obtained through a shared encoder. The Key contributions include: (i) demonstrating how a single set of common features can be extracted from an I/Q dataset for training multiple downstream RF classification or regression tasks, (ii) designing a method of structured noise infusion that impairs the reverse recovery of the raw I/Q from the feature space without significant impact to the downstream tasks, and (iii) demonstrating the application of this method on informative real-world, publicly available datasets. Sharing well-constructed yet privacy-preserving features can accelerate foundational research on ML for wireless without relying on hard-to-get I/Q samples.
TrustSense: Structure-Adaptive Trust Inference for Resilient Cooperative IoT Sensing Ruobing Li, Yan (Ian) Zhang (Arizona State University); Yan Zhang (University of Akron); Rui Zhang (University of Delaware); Yanchao Zhang (Arizona State University) 11:19 - 11:32
Abstract: Cooperative IoT sensing is vulnerable to noise, hardware failures, adversarial manipulation, and dynamic participation—factors that degrade reliability beyond static, device-level trust models. We present TrustSense, an inductive, structure-adaptive framework for scalable per-report trust evaluation. During training, TrustSense learns a compact subgraph via dynamic edge masking, suppressing noisy and adversarial connections while preserving spatial consistency. A forced anchor-attention mechanism stabilizes message propagation by grounding it to historically reliable devices, ensuring robustness under drift and attacks. At runtime, a three-stage pipeline—historical consistency initialization, iterative trust scoring, and targeted fine-tuning for new devices—enables efficient, cold-start-aware evaluation with low overhead. Experiments on real and synthetic datasets show that TrustSense consistently outperforms state-of-the-art baselines in accuracy, robustness, and scalability across diverse attack scenarios and participation patterns. On the large-scale LoED dataset, TrustSense improves detection by over 4.7 times and calibration by two orders of magnitude over baselines. It sustains approximately 90% accuracy and latency below 250 ms across thousands of devices, tolerating 50% malicious nodes.
CEDAR: Structure-Aware Contrastive Learning for One-Class Network Intrusion Detection Yiming Chen, Erfan Meskar, Ben Liang (University of Toronto) 11:32 - 11:45
Abstract: Intrusion detection remains a critical challenge in modern networks. Existing approaches primarily attempt to detect attacks by identifying anomalies in network flow features, which is significantly challenged by the high cost of labeling, the limited availability of attack data, and the heterogeneous and continually evolving nature of attack techniques. In this paper, we propose CEDAR, a one-class feature-based detector that requires only benign traffic data in training, by integrating weakly structured contrastive representation learning with von Mises-Fisher mixture modeling in the latent space. CEDAR leverages communication context derived from source-destination interactions as a form of weak supervision to learn structure-aware embeddings, eliminating the need for synthetic flow augmentations. Furthermore, it jointly learns the encoder and the von Mises-Fisher mixture density parameters under a unified objective with proven convergence guarantees. Comprehensive experiments on multiple benchmark datasets demonstrate that the proposed framework achieves favorable performance compared with a diverse set of intrusion detection baselines.
Skynet: Workflow-Level Anomaly Detection for Agentic AI via Semantic and Structural Modeling Chaoyu Zhang, Hexuan Yu, Heng Jin (Virginia Tech); Shanghao Shi, Ning Zhang (Washington University in St. Louis); Yi Shi, Yulia Gel, Y. Thomas Hou, Wenjing Lou (Virginia Tech) 11:45 - 11:58
Abstract: Agentic AI systems execute complex tasks through long-horizon workflows of planning, tool use, and multi-agent coordination. Task failures in these systems often originate from a single step, such as an injected prompt or a flawed plan, and are then amplified through downstream dependencies as the corrupted step propagates across many subsequent agents and tool calls. Existing defenses either target a specific class of attacks or failures, or inspect individual prompts and steps in isolation. Both leave the global dependency structure of a workflow unexamined, and miss the inconsistencies that only emerge when the execution is viewed as a whole. We argue that anomaly detection for agentic AI must reason at the workflow level, where global execution structure exposes signals that local checks cannot see. We present Skynet, a principled workflow-level anomaly detection framework that turns observed multi-agent execution into directed workflow graphs and scores them against learned benign behavior. Skynet jointly models the semantic execution context and the structural organization of inter-agent delegation, tool invocation, and data-flow dependencies, and trains only on benign workflows. Because training never sees attacks or failures, this design naturally extends to zero-day detection: any execution that violates benign workflow regularities surfaces as off-manifold geometry under a single decision rule. We evaluate Skynet on three public agentic safety and failure benchmarks. It sustains high recall together with a sub-1% false positive rate, with per-workflow and per-step latencies low enough for online monitoring of agentic AI runtimes.
11:58 - 12:00 Session Wrap-Up
12:00 - 13:30 Break
13:30 - 15:00 Poster and Demo Session
Poster/demo titles and presenters TBD
15:00 - 15:25 Coffee Break
15:25 - 16:45    Session 10: IoT Systems: Sensing, Interaction, and Communication
Session Chair: TBD
Towards Ubiquitous Backscatter Capacitive Sensing Ningzhi Zhu, Rong Ding, Zijie Chen, Yi Fu (Shanghai Jiao Tong University); Guiyun Fan (Tongji University); Chenhao Ying, Jiale Lei (Shanghai Jiao Tong University); Yanhong Jiao, Xiangchen Ma (China Mobile Group Design Institute); Haiming Jin (Shanghai Jiao Tong University) 15:25 - 15:38
Abstract: This paper presents CapScatter, the first sub-10$\mu W$ backscatter capacitive sensor compatible with sensing capacitors in 10-1000$pF$, marking a step toward ubiquitous backscatter capacitive sensing. Built on three key designs, we overcome the limitation of existing sub-10$\mu W$ works compatible only with sensing capacitors below 10$pF$. Specifically, (i) we design an excitation-driven capacitance encoding circuit, which consists of decoupled capacitance measurement and backscatter communication subcircuits, to cover the wide capacitance range; (ii) we design a power-scanning capacitance decoding algorithm, in which the reader queries the sensor multiple times at different transmitted powers, to eliminate the interference caused by the received excitation power variations; (iii) we design a collision avoidance circuit that adds only a few passive components with negligible additional power consumption, to avoid signal collision between multiple sensors. Our experiments show that CapScatter achieves a median measurement error of 2.17$\%$ and a maximum measurement error of 3.85$\%$ over a capacitance range of 10-1000$pF$. Meanwhile, its static power is 4.32$\mu W$, and the operating power ranges from 6.11$\mu W$ to 20.51$\mu W$.
Democratizing LoRa Weak Signal Demodulation Zijie Chen, Haiming Jin, Rong Ding, Ningzhi Zhu (Shanghai Jiao Tong University); Guiyun Fan (Tongji University); Kai Ying, Chenhao Ying (Shanghai Jiao Tong University); Dongming Shan (Shanghai Kupas Technology Co., Ltd); Cheng Wang (Tongji University); Lu Su (Purdue University); Linghe Kong (Shanghai Jiao Tong University, China) 15:38 - 15:51
Abstract: This paper democratizes LoRa weak signal demodulation from both the methodological and experimental perspectives. Methodologically, by drawing inspiration from LoRa’s CSS modulation principle, this paper proposes DeLoRa, the first lightweight neural-enhanced method for LoRa weak signal demodulation which can be executed on LoRa nodes possessing far more limited hardware resources than LoRa gateways. Experimentally, this paper proposes MCU-SDR, the first hardware testbed that matches the stringent resource constraints of LoRa nodes, enabling LoRa researchers to carry out in-situ evaluation of demodulation methods for LoRa nodes. Our extensive evaluation in indoor and outdoor environments shows that DeLoRa operates effectively (i.e., ensures <10% symbol error rate) at SNR level as low as -18.8 dB, achieving a \textbf{3.3 dB SNR gain over SOTA demodulation methods that can run on LoRa nodes}.
FSMA: Scaling LoRa Satellite IoT with Link-aware Multiple Access Rohith Reddy Vennam (University of California San Diego); Maiyun Zhang (University of California, San Diego); Raghav Subbaraman (University of California San Diego); Deepak Vasisht (University of Illinois Urbana-Champaign); Dinesh Bharadia (UCSD) 15:51 - 16:04
Abstract: LoRa-based non-terrestrial networks are rapidly becoming infrastructure for global IoT, with LEO satellites delivering direct-to-device connectivity and drone-mounted gateways serving precision agriculture, emergency response, and environmental monitoring. Both settings share a fundamental MAC-layer bottleneck: large coverage footprints cause thousands of devices to collide on the same channel, while gateway mobility induces rapid link variations, making link-unaware transmissions a persistent source of wasted airtime. These challenges demand random access solutions, yet ALOHA provides no collision coordination, CSMA sensing fails beyond ~15 km, and BSMA's continuous busy-tone consumes 80-90\% of channel capacity. We propose Free Signal Multiple Access (\textit{FSMA}), a synchronization-free, gateway-controlled MAC protocol. Like a traffic signal, \textit{FSMA} uses a single LoRa upchirp (\textit{FreeChirp}) to indicate channel availability. By transmitting it at a lower spreading factor than the node uplink, \textit{FSMA} implicitly filters weak-link nodes through channel reciprocity, reducing the collision window by 6-200$\times$ without synchronization, handshakes, or hardware modifications. Hardware experiments with 25 COTS LoRa nodes and a drone-mounted gateway demonstrate 2$\times$ higher throughput, 2.5-5$\times$ better packet reception ratio, and up to 5$\times$ improved energy efficiency. Large-scale simulations with real TLE-based satellite trajectories confirm scalability to 5000+ devices per satellite pass.
BrewSense:A Mobile Water quality monitoring platform using laser reflection near Brewster angle Qianfei Ren, Juntao Zhou, Longyuan Ge, Zeyuan Ding, Dian Ding, Yida Wang, Zhuxi Chen (Shanghai Jiao Tong University); Yijie Li (National University of Singapore); Hao Pan, Jiadi Yu, Yi-Chao Chen (Shanghai Jiao Tong University); Lili Qiu (Microsoft Research Asia, UT Austin); Guangtao Xue (Shanghai Jiao Tong University) 16:04 - 16:17
Abstract: Water quality monitoring has long been a crucial component of environmental conservation efforts. In this paper, we present a mobile platform-based system employing specialized laser emission and optical reception path configurations to collect reflective light spot information near the Brewster’s angle in aqueous environments. By calculating the refractive index corresponding to the Brewster’s angle, our system enables real-time detection of water quality anomalies. This solution achieves active, non-invasive monitoring through a mobile deployment platform, demonstrating a detection accuracy of 1g/L in solution concentration measurements.
ECO-COMM: An Ultra Low-Latency Event Camera based Optical Communication System Chengling Xu (University of Wisconsin-Madison); Keigo Hirakawa (University of Dayton); Feng Ye (University of Wisconsin-Madison) 16:17 - 16:30
Abstract: Ultra‑low‑latency communication is a key requirement for emerging Next‑generation (NextG) applications such as immersive XR, real‑time collaborative control, and distributed sensing. This paper presents ECO‑COMM, an event‑camera‑based optical communication system designed to enable ultra low‑latency device association and lightweight information exchange. ECO‑COMM leverages the asynchronous sensing and microsecond ($\mu$s)‑level temporal resolution of event cameras to capture high‑frequency spatiotemporal light modulation, eliminating the latency overhead of frame‑based acquisition, handshake procedures, and centralized scheduling. We systematically analyze hardware‑induced challenges, including timestamp inconsistency, readout contention, trailing effects, and inevitable refractory period, and introduce targeted mitigation techniques to ensure robust operation. A complete prototype implementation using an eight‑LED transmitter and an off‑the‑shelf event camera demonstrates 10~$\mu$s-level device association and 100~$\mu$s-level symbol latency. Experimental results further show end‑to‑end transmission latency below 8~ms for 32‑byte payloads at 0.1\% bit error rates. ECO‑COMM establishes a practical and complementary communication paradigm for ultra‑low‑latency NextG systems where responsiveness and temporal precision are paramount.
IoTell: Multimodal Sensor Fusion for Enhanced IoT Control Md Fazlay Rabbi Masum Billah, Amod K. Agrawal (Amazon Lab126); Viswajith Govinda Rajan, Brad Campbell (University of Virginia) 16:30 - 16:43
Abstract: Current control interfaces between a user and a network of heterogeneous IoT devices largely consist of device-specific applications or voice commands with a fixed set of instructions for each device. Such control mechanisms fail to cope with the vision of dense trillion-node IoT network, where users would interact with tens of smart devices in everyday environments. Recent advances in RF-based ranging technologies enable accurate estimation of user location in indoor environments. However, leveraging this spatial context to enable intuitive interaction remains an open challenge, particularly in dense deployments where multiple candidate devices exist within close proximity. In this paper, we introduce IoTell, an intuitive control interface that allows users to interact with a device by pointing to it and issuing simple voice commands. IoTell fuses location, orientation, and voice instructions to identify the intended device without requiring explicit device naming. Operating in device-dependent mode, the system estimates user location and orientation through a wearable device using RF-based localization and inertial sensing. In the device-independent mode, IoTell enables the user to have a similar experience even when the user may not carry a wearable, by estimating user location and orientation through multimodal fusion of audio and RF signal features. We implement IoTell using a UWB-based localization system and inertial sensing on a wearable device, and evaluate its performance in a multi-device indoor environment. Our experiments demonstrate that IoTell achieves 95.2\% accuracy in correctly executing commands in device-dependent settings and 88.3\% accuracy, with a mean localization error of 45 cm, in device-independent scenarios. These results show that multimodal fusion of spatial and linguistic context can significantly improve interaction with IoT systems under realistic conditions.
16:43 - 16:45 Session Wrap-Up
16:45 - 17:00 Concluding Remarks

Thursday, November 26, 2026

09:00 - 17:00 Workshops
Workshop-AI-RAN
Workshop-AIoT
Workshop-FIDTA
Workshop-CyberNet
Workshop-ML4NxtGNet
Workshop-Beyond AoI

Banner image credits: David Edelstein, Denys Nevozhai, and Elle Edwards under the Upsplash licence.