INFOCOM 2026 Papers

A key bottleneck in federated learning (FL) is the high communication cost of transmitting large local model updates. This paper proposes Entropy-Adaptive FL (EA-FL), a novel framework that integrates information-theoretic coding into the FL pipeline to reduce the bit rate of encoded updates. In EA-FL, clients train local models under an entropy constraint on their updates, ensuring they are compressible and require fewer bits when entropy coded. Accounting for wireless channel imperfections, we analyze the statistical behavior of the decoded updates at the server and derive bounds on the first and second derivatives of the EA-FL loss function, establishing its Lipschitz continuity. These results enable a convergence analysis that explicitly captures the effects of quantization error, channel noise, and aggregation. Building on this, we develop an optimization framework for efficient bit allocation across clients under a fixed total quantization budget. Extensive experiments show that (i) EA-FL outperforms state-of-the-art quantized FL methods in rateaccuracy trade-offs, and (ii) the proposed bit allocation scheme significantly improves over uniform allocation under the same bit budget.

Industrial Deep Learning Recommendation Models (DLRMs) comprise memory-intensive embedding operations and compute-intensive DNN layers, often resulting in suboptimal GPU resource utilization under high-throughput inference workloads. However, the memory demands of DNN layers vary considerably across different workloads and execution phases, often leading to unpredictable interference, which makes it challenging to efficiently co-execute embedding and DNN operators.This paper presents RecFlow, a high-performance DLRM serving framework that leverages intra-SM parallelism to co-run embedding and DNN computations through fine-grained resource coordination. RecFlow profiles the workload characteristics of each DNN phase and applies adaptive parallel strategies to sustain high memory bandwidth utilization while minimizing interoperator interference. To further enable parallelism between the structurally dependent embedding and top-DNN stages, RecFlow introduces an incremental batching mechanism that overlaps their execution using newly arrived requests, thereby enabling inter-batch parallelism without incurring additional queuing latency. Extensive evaluations on real-world production workloads demonstrate that RecFlow improves serving throughput by up to 1.13 × over state-of-the-art DLRM inference systems while reducing latency in high-throughput serving scenarios.

In a Hierarchical Inference (HI) system, an end device processes data samples using a local Deep Learning (DL) model. Only when the confidence of the local DL is less than a chosen threshold, it offloads the sample for remote DL inference on an edge server. By balancing the misclassifications and offloading costs, a HI system can improve accuracy and responsiveness and reduce bandwidth usage. To this end, prior work studied a HI learning problem for classification tasks, aiming to learn an optimal threshold on the confidence to minimize the cumulative costs. However, existing formulations considered (i) abstract offloading costs and (ii) assumed that the remote DL inference for an offloaded task is available by the end of each round. In contrast, considering that end devices are wirelessly connected to edge servers, we study a HI system with an offload queue where the offloading cost is implicitly modeled by the queue length. Also, we consider the practical setting, where the remote DL inference is obtained only after the task is served from the offload queue. For bounded but unknown arrival and service processes, we formulate the problem of average mismatch error minimization subject to queue stability. We propose a BOLD Hedge-Q policy based on Lyapunov optimization. The novelty of BOLD Hedge-Q lies in solving a delayed-feedback online learning problem — with sub-linear regret — that results from the Lyapunov drift-plus-penalty analysis. Given a penalty parameter V ≥ 1, we show that under BOLD Hedge-Q, the queue is upper bounded by V + λm, and the mismatch error is within an additive factor of O(1/V) from that of the optimal policy, where λm is the maximum number of arrivals in a round.

Edge-cloud systems are increasingly deployed to meet the demands of real-time applications by processing latency-sensitive tasks at the edge site, near the end-devices. When load exceeds capacity, offloading of requests to the cloud allows the system to balance the spike. As their popularity increases, edge servers become an attractive target for attackers whose goal is to degrade system performance.This paper investigates worst-case (i.e., damage-maximizing) attacks on reactive edge-cloud systems, focusing on two key performance metrics: expected end-to-end latency and the probability of violating service-level agreements (SLAs). We use a general modeling framework based on two layers of computation and analyze how adversarial removal of edge servers affects delay under reactive offloading mitigation.We adopt common queueing models and prove that the system’s performance remain convex in the attack size, even when accounting for reactive offloading by the edge. This enables a characterization of worst-case policies as concentrated attacks, targeting only a few critical sites to maximize damage, yielding greater performance degradation than spreading efforts across the network. Simulation results validate the analysis and demonstrate that the identified worst-case attack strategies can increase damage by up to 40% using the same attack resources.

In this paper we demonstrate a novel downgrade attack for 5G networks able to dramatically reduce the efficiency of Multi-Input-Multi-Output links, while remaining virtually invisible to standard intrusion monitors.By opportunistically injecting noise precisely aligned with silent reference signals in the radio frames (called Zero-Power Channel State Information Reference Signals), we show that an attacker can corrupt the feedback mechanism implemented at the User Equipment for channel estimation, forcing the base station to downgrade or disable spatial multiplexing. This mechanism is much simpler than pilot spoofing schemes and dramatically harder to defend: indeed, silent reference signals cannot be protected by integrity mechanisms and their sparsity keeps the average power of the injected noise difficult to detect.We design a robust mechanism for synchronizing the attacker transmissions to the silent reference signals, and quantify the latency of the attack under different SNR and cell load conditions. The approach has been implemented in a Software Defined Radio testbed and experimentally validated in a private network, by demonstrating a success probability higher than 82% on two different commercial smartphones and a throughput degradation up to 70%.

Blood pressure (BP) is a critical indicator for assessing cardiovascular health. However, existing wearable and Radio Frequency (RF)-based methods typically require users to maintain static postures to obtain stable measurements, failing to ensure continuity during daily activities. To address this, we propose MuBP, which achieves continuous BP measurement via multimodal data fusion of Ultra-Wideband (UWB) and Inertial Measurement Unit (IMU) sensors in commercial smartwatches. The MuBP system first extracts Channel Impulse Responses (CIR) from UWB signals and constructs a motion-aware state matrix to effectively separate wrist motion artifacts from pulse signals. Subsequently, a periodic feature extraction network is designed to capture multi-scale pulse waveform features. Finally, a Bidirectional Long Short-Term Memory (Bi-LSTM) network is adopted to model temporal dependencies, combined with attention mechanisms for adaptive feature selection. We implement the MuBP system on a commercial mobile device and conduct 500 experiments involving 50 participants. Extensive results show that MuBP achieves Mean Errors (ME) and Standard Deviations (STD) of 1.86 ± 6.16 mmHg for systolic BP and 1.08 ± 5.6 mmHg for diastolic BP, maintaining stable measurements under diverse factors and complying with the Association for the Advancement of Medical Instrumentation (AAMI) SP10 standard. Additionally, MuBP supports real-time hypertensive risk alerts, providing an efficient solution for continuous non-invasive BP monitoring in daily life.

Existing fuzzing tools struggle to analyze proprietary Industrial Communication Protocols (ICPs) due to the lack of detailed protocol specifications. Consequently, they often rely on manual analysis or coarse-grained black-box approaches, which offer limited insight into protocol field semantics. This ultimately restricts both the coverage and depth of the testing process. To address these limitations, we propose a novel fuzzing framework called ICProFuzz, which leverages Large Language Models (LLMs) to understand proprietary ICPs. ICProFuzz utilizes LLMs to deeply analyze captured communication messages and available partial protocol descriptions. This enables the automated extraction of field semantics, length constraints, and inter-field dependencies. Utilizing the inferred field types and syntactic formats, we design a suite of advanced mutation strategies, including boundary value injection, format-preserving assignment, and protocol-aware keyword insertion. Finally, the generated test cases are fed into the fuzzing engine, where both the semantic parsing and mutation strategies are dynamically adjusted based on test coverage and anomaly feedback. We conducted extensive evaluations using three widely used proprietary ICPs: S7comm, UMAS, and CIP. The experimental results show that ICProFuzz outperforms traditional baseline fuzzers in vulnerability identification, path coverage, and branch coverage. Moreover, ICProFuzz identified 14 vulnerabilities across six real devices, including two zero-day vulnerabilities.

The rapid growth of low Earth orbit (LEO) satellite constellations has revolutionized broadband access, Earth observation, and direct-to-device connectivity. However, the expansion of ground station infrastructure has not kept pace, creating a critical bottleneck in satellite-to-ground backhaul capacity. Traditional parabolic dish antennas, though effective for geostationary (GEO) satellites, are ill-suited for dense, fast-moving LEO networks due to mechanical steering delays and their inability to track multiple satellites simultaneously. Phased array antennas offer electronically steerable beams and multi-satellite support. However, their integration into ground stations is limited by the high cost, hardware issues, and complexity of achieving sufficient antenna gain. We introduce ArrayLink, a distributed phased array architecture that coherently combines multiple small, commercially available panels to achieve high-gain beamforming and unlock line-of-sight MIMO spatial multiplexing with minimal additional capital expenditure. By spacing 16 32×32 panels across a kilometer-scale aperture, ArrayLink enters the radiative near-field, focusing energy in both angle and range while supporting up to four simultaneous spatial streams on a single feeder link. Through rigorous theoretical analysis, detailed 2D beam pattern simulations and real-world hardware experiments, we show that ArrayLink (i) achieves dish-class gain exceeding that of a 1.47 m reflector, (ii) maintains four parallel streams at ranges of hundreds of kilometers (falling to two beyond 2000 km), and (iii) exhibits tight agreement across theory, simulation, and experiment with minimal variance. These findings pave the way for a practical and scalable approach to boosting satellite backhaul capacity.

Low Earth Orbit (LEO) satellite networks have attracted considerable attention for their ability to deliver global, low-latency broadband Internet services. In this paper, we present a large-scale measurement study of the Starlink network, the largest LEO satellite constellation to date. We first propose an efficient method for discovering active Starlink user routers, identifying approximately 5.98 million IPv6 addresses across 208 regions in 165 countries. Compared to general-purpose IPv6 target generation algorithms, our router-centric approach achieves near-complete coverage and, to the best of our knowledge, yields the most comprehensive known set of active IPv6 addresses for Starlink user routers. Based on the discovered user routers, we further propose an efficient method for mapping the Starlink backbone network and uncover a topology consisting of 49 Points of Presence (PoPs) interconnected by 98 links. We conduct a detailed statistical analysis of active Starlink user routers and PoPs, and further characterize the IPv6 address assignment strategy adopted by the Starlink network. Finally, we analyze the latency of Starlink user routers, propose a method to distinguish different types of users within the same region using outside-in measurement, and identify the ongoing V2 Mini satellite deployment as a potential driver of the performance improvements. The dataset of the Starlink backbone network is publicly available at https://ki3.org.cn/#/starlink-network.

Keystroke inference attacks severely threaten data security and user privacy. Existing methods often exploit typing-induced signals—such as vibrations, acoustics, and visual cues—but typically require close proximity, line-of-sight, or precise transceiver placement, limiting their practicality. We propose TagStroke, a non-invasive, wireless keystroke inference attack that leverages low-cost COTS RFID systems. TagStroke uses passive UHF RFID tags placed beneath keyboards and a concealed RFID reader to detect typing-induced signal changes, enabling accurate keystroke recognition and semantic recovery.TagStroke addresses key challenges, including RFID signal instability, interference from hand and body movements, and limited spatial sensing resolution. To overcome these, it introduces robust signal preprocessing, a hybrid Temporal Convolution-Transformer model for keystroke detection, and a multi-stage recognition framework integrating spatial decoding with large language models (LLMs). We prototype and evaluate TagStroke with 11 volunteers using a low-cost COTS RFID setup. TagStroke achieves 97.21% keystroke detection accuracy, 87.24% keystroke recognition accuracy, and a word error rate (WER) of 21.06%, along with high content recovery similarity scores e.g., 0.71 at 2 m. These results rival existing CSI-based attacks while avoiding their constraints, demonstrating TagStroke’s practicality, accuracy, and cost-effectiveness.

Localization is a key component of the wireless ecosystem. Machine learning (ML)-based localization using channel state information (CSI) is one of the most popular methods for achieving high-accuracy localization with low cost. However, to be accurate and robust, ML-based algorithms need to be trained and tested with large amounts of data, covering not only many user equipment (UE)/target locations, but also many different access points (APs) locations to which the UEs connect, in a variety of different environment types. This paper presents a massive-sized CSI dataset, WiLoc (Wi-Fi Localization), and makes it publicly available. WiLoc is obtained by a series of precision measurement campaigns that span three months, and it is massive in all the above-mentioned three dimensions: > 12 million UE locations, > 3,000 APs, covering 16 buildings for indoor localization, and > 30 streets for outdoor use. The paper describes the dataset structure, measurement environments, measurement protocols, and the dataset validations. Comprehensive case studies validate the advantages of large datasets in ML-driven localization strategies for both "standard" and transfer learning. We envision this dataset, which is by far the largest of its kind, to become a standard resource for researchers in the field of ML-based localization.

Software packet classifiers are essential components in cloud environments, where they must balance fast rule lookup, rapid rule updates, and minimal memory overhead. Existing approaches—such as decision tree-based and tuple space-based methods—typically prioritize one performance metric at the expense of the others. We propose TurboTSS, a hybrid solution that achieves the fastest rule update speed in tuple-space-based solutions, and incorporates four key enhancements to accelerate rule lookup: selecting per-rule significant fields to build tuples, using tries to filter unnecessary tuples, using cross-field trees to fetch rules within a tuple, and using extra storage to lookup big rules. Our evaluations on the TurboTSS prototype show that TurboTSS provides the fastest rule lookup speed and moderate-fast rule update speed, sacrificing only some extra memory space, compared with state-of-the-art solutions.

Much of federated learning (FL) focuses on settings where local dataset statistics remain the same between training and testing. However, this assumption often does not hold in practice due to distribution shifts, motivating the development of domain generalization (DG) approaches that leverage source domain data to train models capable of generalizing to unseen target domains. In this paper, we are motivated by two major gaps in existing work on FL and DG: (1) the lack of formal mathematical analysis of DG objectives; and (2) DG research in FL being limited to the star-topology architecture. We develop Decentralized Federated Domain Generalization with Style Sharing ($\textit{StyleDDG}$), a decentralized DG algorithm which allows devices in a peer-to-peer network to achieve DG based on sharing style information inferred from their datasets. Additionally, we provide the first systematic approach to analyzing style-based DG training in decentralized networks. We cast existing centralized DG algorithms within our framework, and employ their formalisms to model $\textit{StyleDDG}$. We then obtain analytical conditions under which convergence of $\textit{StyleDDG}$ can be guaranteed. Through experiments on popular DG datasets, we demonstrate that $\textit{StyleDDG}$ can obtain significant improvements in accuracy across target domains with minimal communication overhead compared to baseline decentralized gradient methods.

Recent advancements in Large Language Models (LLMs) have transformed many fields including scientific discovery, content generation, biomedical text mining, and educational technology. However, the substantial requirements for training data, computational resources, and energy consumption pose significant challenges for their practical deployment. A promising alternative is to leverage smaller, specialized language models and aggregate their outputs to improve overall response quality. In this work, we investigate a networked LLM system composed of multiple users, a central task processor, and clusters of topic-specialized LLMs. Each user submits categorical binary (true/false) queries, which are routed by the task processor to a selected cluster of $m$ LLMs. After gathering individual responses, the processor returns a final aggregated answer to the user. We characterize both the information accuracy and response timeliness in this setting, and formulate a joint optimization problem to balance these two competing objectives. Our extensive simulations demonstrate that the aggregated responses consistently achieve higher accuracy than those of individual LLMs. Notably, this improvement is more significant when the participating LLMs exhibit similar standalone performance.

We consider the design of mixing matrices to minimize the operation cost for decentralized federated learning (DFL) in wireless networks, with focus on minimizing the maximum per-node energy consumption. As a critical hyperparameter for DFL, the mixing matrix controls both the convergence rate and the needs of agent-to-agent communications, and has thus been studied extensively. However, existing designs mostly focused on minimizing the communication time, leaving open the per-node energy consumption that is critical for energy-constrained devices. This work addresses this gap through a theoretically-justified solution for mixing matrix design that aims at minimizing the maximum per-node energy consumption until convergence, while taking into account the broadcast nature of wireless communications. Based on a novel convergence theorem that allows arbitrarily time-varying mixing matrices, we propose a multi-phase design framework that activates time-varying communication topologies under optimized budgets to trade off the per-iteration energy consumption and the convergence rate while balancing the energy consumption across nodes. Our evaluations based on real data have validated the efficacy of the proposed solution in combining the low energy consumption of sparse mixing matrices and the fast convergence of dense mixing matrices.

Accurate modeling of urban-scale wireless channels is critical for the next-generation 5G/6G communication systems. Conventional approaches, both deterministic and statistical are inherently limited by a trade-off between accuracy and computational efficiency. While recent advances in neural radiance fields (NeRF2) have shown promise for indoor channel prediction, their extension to urban-scale environments remains infeasible due to the need for large-scale labeled measurements and the substantial computational overhead required for high-resolution radiance field representations. To address these challenges, we propose SIGN-RF, a self-adaptive neural field framework for urban-scale radio reconstruction. SIGN-RF introduces two key innovations: (1) adaptive hierarchical sampling that autonomously identifies electromagnetically significant regions without environmental priors, and (2) self-supervised position labeling that efficiently associates unpositioned RF measurements with spatial coordinates, thereby maximizing the utility of heterogeneous measurement streams. Comprehensive real-world experiments show that SIGN-RF achieves state-of-the-art accuracy in large-scale outdoor radio reconstruction, converging 67% faster than existing NeRF2, while requiring as few as 20% positioned measurements. These results underscore the scalability and data efficiency of SIGN-RF for urban wireless channel modeling.

5G mobile networks are becoming new targets for Advanced Persistent Threat (APT) attacks. While existing provenance-based intrusion detection systems (PIDS) show promise for APT detection, they are inadequate for complex 5G networks due to the absence of explainable investigation results. They generate alert graphs with numerous false positive nodes, imposing substantial cognitive burdens on security analysts. This paper presents SAGE, the first end-to-end framework for automated APT investigation in 5G networks. SAGE presents a two-stage approach. For improved detection precision, SAGE constructs global heterogeneous provenance graphs integrating system and 5G application logs with semantic embeddings for efficient node-level anomaly detection. For automated investigation, SAGE introduces a novel self-reflective and self-reasoning LLM framework with domain-specific Retrieval Augmented Generation (RAG). Through carefully designed APT investigation workflows, it reduces detection false positives and automatically generates comprehensive natural language reports explaining attack tactics, techniques and procedures (TTPs) and impacts, bridging the critical gap between alerts and human-friendly intelligence. Extensive experiments on the constructed 5G APT dataset demonstrate that SAGE’s investigation framework improves node-level detection precision by an average of 12% and generates high-quality investigation reports, achieving superior TTP-level detection performance with over 80% precision, significantly outperforming state-of-the-art methods.

RFID offers a promising foundation for sensing due to its low cost, battery-free operation, and wide commercial deployment. However, sensing RF-insensitive parameters typically requires specialized circuits or hardware modifications to RFID tags, limiting scalability and practical adoption. In this paper, we present PicoTag, a novel design paradigm that seamlessly transforms commodity RFID tags into versatile sensing nodes capable of detecting RF-insensitive environmental parameters. Implemented as a low-cost, low-power add-on patch, PicoTag leverages electromagnetic coupling effects to operate alongside unmodified RFID tags. Specifically, we design a new framework capable of integrating impedance-sensitive media from prior work, thereby enabling generalized sensing. For robust parameter reading, we further introduce a passive reference-signal generation circuit that enables calibration, compensating for link coefficients and hardware variations. Importantly, PicoTag is fully compatible with COTS RFID communication protocols and introduces no interference to existing RFID systems. We develop a complete hardware prototype of PicoTag and evaluate its performance across a range of sensing tasks and deployment scenarios. The results show that the power consumption of PicoTag is 15µW, achieving an average temperature sensing error of just 0.9°C while maintaining consistent performance across different tags, deployment environments, and distances. The hardware schematics can be found at: https://doi.org/10.5281/zenodo.18085137.

As applications with high bandwidth and computation are increasingly deployed in data centers, host congestion emerges with significant packet drops and tail latency increases, which is on account of the contention on memory bandwidth between the NIC and CPU. In this paper, we propose a precise approach to mitigate the host congestion by offloading partial mitigation logic to DPU to significantly avoid packet loss and tail latency inflation. Our primary idea is to collect host congestion signals and perform corresponding precise congestion mitigation strategies in DPU. To this end, we first collect the signals in the Linux kernel reflecting different congestion levels, and share them with DPU through the memory sharing mechanism. The mitigation mechanism in DPU is then designed with three ideas, i.e., latency-sensitive packets giving priority on transmission, early precise congestion notification by the ECN label, and active packet drop management. Finally, the proposed approach is evaluated by our prototype implementation on NVIDIA BlueField-3 commodity DPU. The experimental results indicate that compared to HostCC, it respectively reduces the packet loss by up to 65.5% and the tail latency by up to 34.9% in the mitigation process with low overheads.