INFOCOM 2026 Papers

IPv6 target generation is essential for conducting rapid Internet-wide scans of IPv6 network assets. Due to the under-utilization of the seed address source and insufficient exploration of seed address structure, existing IPv6 target generation algorithms suffer from limited active address prediction scale and prolonged prediction time, hampering the ability to predict billions of IPv6 active addresses in a few hours. An efficient IPv6 large-scale target generation framework, 6Massive, is proposed. 6Massive introduces a sampling-based IPv6 seed address expansion strategy to improve the utilization rate of the seed address source, a merged divisive hierarchical clustering strategy to fully leverage seed address structural information and expand the low-dimensional IPv6 address pattern space, and a feedback strategy to identify active high-dimensional IPv6 address patterns without the need for extra pre-scanning. Compared to 5 typical algorithms (6Gen, 6Tree, 6Hit, 6Scan, and HMap6), 6Massive demonstrates superior performance across 4 seed address sets (the number ranges from 100 thousand to 1 million). Specifically, 6Massive can probe 156.84% to 533.94% more IPv6 active addresses than these 5 algorithms without the feedback strategy and 279.03% to 484.16% more with it. Within 1.76 days, 6Massive can predict 1.644 billion IPv6 active addresses.

Low Earth Orbit (LEO) satellite communication is a promising approach to providing Internet connectivity to users in many remote areas. As videos are likely to account for most traffic in the LEO satellite network, as in the rest of the Internet, this work introduces a novel video-aware mobility management framework tailored for LEO satellite networks. Utilizing simulation models alongside real-world datasets, we show the importance of handoff strategy and throughput prediction algorithms in single-user and multi-user video streaming scenarios. Motivated by these observations, we propose a set of novel algorithms that can jointly choose the satellite and video bitrate to optimize the Quality of Experience (QoE). We first develop Model Predictive Control (MPC) and Reinforcement Learning (RL) based algorithms for a single user, and then extend them to accommodate multiple competing users that may share the same satellite. We introduce centralized training and distributed inference for our RL design, enabling a distributed policy informed by a global perspective. We demonstrate the effectiveness of our proposed models using trace-driven simulation and testbed experiments. We share our code and data with the research community.

The deployment of large language models' (LLMs) inference at the edge can facilitate prompt service responsiveness while protecting user privacy. However, it is critically challenged by the resource constraints of a single edge node. Distributed inference has emerged to aggregate and leverage computational resources across multiple devices. Yet, existing methods typically require strict synchronization, which is often infeasible due to the unreliable network conditions. In this paper, we propose HALO , a novel framework that can boost the distributed LLM inference in lossy edge network. The core idea is to enable a relaxed yet effective synchronization by strategically allocating less critical neuron groups to unstable devices, thus avoiding the excessive waiting time incurred by delayed packets. HALO introduces three key mechanisms: (1) a semantic-aware predictor to assess the significance of neuron groups prior to activation. (2) a parallel execution scheme of neuron group loading during the model inference. (3) a load-balancing scheduler that efficiently orchestrates multiple devices with heterogeneous resources. Experimental results from a Raspberry Pi cluster demonstrate that HALO achieves a 3.41x end-to-end speedup for LLaMA-series LLMs under unreliable network conditions. It maintains performance comparable to optimal conditions and significantly outperforms the state-of-the-art in various scenarios.

Deploying large language models (LLMs) on edge devices often demands effective fine-tuning to adapt to local requirements, yet confronts two key challenges: data heterogeneity and resource constraints. Existing solutions fall short: traditional federated learning methods, inherently designed for small-scale models, fail to scale to the size and complexity of LLMs. Meanwhile, current federated fine-tuning approaches for LLMs fail to adequately address the stringent resource constraints of edge devices. To address these, we propose FedDynMask——a framework designed specifically for efficient federated fine-tuning of LLMs on edge devices. FedDynMask couples Dynamic Personalized Sparse Masking (DPSM) on edge devices with Masked Intersection Aggregation (MIA) on the server via dynamic mask updates. DPSM enables devices to dynamically identify and retain only task-essential parameters by leveraging local data distributions, reducing computational overhead and enabling local adaptation. MIA further reduces communication overhead by identifying consistent parameters through masked intersection, preventing global model performance degradation caused by heterogeneous data. Compared to baselines, this design reduces communication overhead to 0.366M and 0.522M for Llama3.2-3B and Llama3-8B models, respectively, and computational overhead to 3.67M and 5.24M. It also achieves a 1.36% accuracy improvement over other methods, making federated fine-tuning of LLMs feasible on resource-constrained edge devices.

As wireless communication systems evolve toward beyond fifth-generation (B5G) and sixth-generation (6G), achieving ultra-high data rates and accommodating massive connectivity present critical challenges for the development of wireless technologies. To meet these challenges, it is important to develop low-cost and high-efficiency multiplexing mechanisms to serve a large number of mobile subscribers with concurrent data streams. We propose a novel scheme, called Power and Polarity Coding Assisted Multiplexing (P2C-MUX), to enhance communication efficiency by leveraging existing signals to enable additional data transmissions, without increasing power and antenna needs. P2C-MUX first recognizes that the interactions of wireless signals can produce an effective signal for data transmission and then utilizes existing communication signals (ECSs) to achieve additional data transmissions. Specifically, a transmitter (Tx) serves its subscriber (i.e., original receiver (Rx)) with multiple ECSs, while simultaneously transmitting to an additional Rx using P2C-MUX. To achieve this, the Tx adjusts the transmit power and polarity of the ECSs such that these modified/coded ECSs can superimpose to generate desired signals at the additional Rx. In essence, P2C-MUX leverages the interactions among ECSs to enable data transmissions for the additional Rx while maintaining the communication performance of the original Rx, thus enhancing multiplexing efficiency. Furthermore, we extend P2C-MUX to more generalized scenarios involving arbitrary numbers of ECSs, additional users, and additional transmissions per user, as well as various modulation schemes, thus broadening its applicability. We utilize (1) the Universal Software Radio Peripheral (USRP) platform for experimental validation of P2C-MUX’s feasibility, and (2) comprehensive MATLAB simulations to demonstrate its effectiveness in improving system spectral efficiency (SE).

Sparse Mixture of Experts (SMoE) has become a preferred architecture for scaling Transformer capacity without increasing computational cost, as it activates only a small subset of experts for each input. However, deploying such an approach for online inference remains challenging due to the large size of a full SMoE model and the complexity of expert routing, especially in resource-constrained edge networks. Moreover, during the online inference, task information is often unavailable, making the task-level routing error-prone. In this work, we propose a novel tree-structured adaptive neural bandit router, Tanbr, to enable efficient and reliable online MoE inference. Instead of relying on explicit task tags, Tanbr estimates the task distribution over time from historical data and uses it to guide task-aware expert merging within a given pre-trained MoE. To handle the large continuous space of merging weights, Tanbr employs a binary tree to progressively partition the space and generate finer candidate weights. It then applies a neural bandit to learn the non-linear mapping from merging weights to model performance and decides optimal expert merging. We prove that Tanbr achieves a sublinear regret bound of $\mathcal{O}(\sqrt T \log (T))$ over T rounds, despite operating over a continuous decision space, matching regret bounds compared to existing methods. Extensive experiments show that Tanbr reduces inference latency by at least 45% and memory usage by up to 25%, while maintaining a high accuracy compared to many state-of-the-art methods.

In many real-time mission-critical systems, Age of Information (AoI) is a key metric for data freshness. However, conventional approaches that minimize merely the average AoI often fail to mitigate the risk of rare but catastrophic high-AoI events, which can be fatal in applications like autonomous driving. We address this gap by introducing a risk-aware scheduling framework that minimizes a weighted combination of the mean AoI and the Conditional Age-at-Risk (CAaR), explicitly managing the AoI’s tail distribution. We formulate the problem as a fractional Markov Decision Process (MDP) and develop RASA (Risk-Aware Scheduling for AoI) algorithm with provable convergence guarantees, which can be further adaptable to prior-free settings. Our theoretical analysis reveals the conditions of the threshold policy, dynamically shifting from "slow but safe" servers to "fast but risky" servers as the current AoI increases. Based on the policy’s threshold structure, we design the Threshold Policy Optimization (TPO) algorithm, a computationally efficient alternative to value iteration. Extensive simulations validate our theoretical results, showing that RASA consistently outperforms optimal risk-oblivious baselines by achieving up to a 32% reduction in risk-aware AoI. We further show the optimality of the threshold-type policy in typical scenarios and demonstrate the proposed schemes’ effectiveness in managing AoI’s tail risk.

Time-sensitive networking (TSN) empowers industrial automation by providing deterministic, low-latency, and low-jitter network communication. Paradoxically, regarding security, such determinism leaks the arrival times of critical traffic (e.g., robotic arm coordination and vehicle brake control), which is exploited to launch side-channel attacks in real-time systems. Considering this, we first identify that TSN is vulnerable to traffic- and memory-based side-channel attacks, and then validate that these attacks can successfully reveal the arrival times of all critical traffic. As a countermeasure, we propose Oblivious TSN (O-TSN), a secure TSN framework that protects the arrival times of critical traffic without over-affecting their performance. We devise two sub-protocols, traffic obfuscation and oblivious priority queues, providing differential privacy and oblivious security, respectively. We formally define the security of O-TSN and ensure strong privacy guarantees by injecting dummy frames while introducing only bounded latency and jitter. We develop a prototype of O-TSN and thoroughly evaluate it on both simulation and Alinx AX7021 boards as real-world hardware. Results show that, compared to standard TSN, our traffic obfuscation forwards all critical traffic in time with 99.5% of original performance, and oblivious queues incur deterministic latency and jitter <0.5ms.

A pretrained neural channel estimator cannot generalize to all channel environments, necessitating online adaptation. Conventional methods demand ground-truth channel coefficients as supervised labels, but such labels are unavailable online. To this end, a self-supervised task is introduced on top of the original channel-estimation task to facilitate label-free adaptation of neural estimators. Specifically, this task randomly masks a fraction of resource elements in each received frame and reconstructs such masked parts. To enable effective reconstruction, the task input must incorporate two components: the unmasked parts and estimated data-symbols of masked parts. These estimated symbols are obtained via an online symbol-recovery mechanism, so no additional pilot overhead is incurred. To consolidate the self-supervised task with the original task, a two-branch masked auto-encoder (MAE) model called ChannelMAE is developed, with each branch dedicated to one task. The two branches share the same encoder but use separate decoders. During online adaptation, the encoder is updated by optimizing the self-supervised branch, which learns channel statistical features and shares them with the channel-estimation branch. Therefore, online channel-estimation accuracy is much improved. Extensive experiments show that ChannelMAE reduces channel-estimation error by up to 71.8% and 87.1% compared with the pretrained model and the state-of-the-art adaptation scheme, respectively.

Constructing a low-Earth orbit (LEO) satellite network is both capital-intensive and time-consuming. To accelerate deployment and reduce costs, inter-constellation sharing, where multiple satellite network operators (SNOs) share their satellite constellations to build LEO networks collaboratively, has emerged as a promising direction. However, the inherent dynamics of LEO satellites inevitably lead to frequent handovers between terrestrial users and core networks operated by different SNOs, posing significant challenges to network continuity and stability.This paper presents Satitch, a new framework that mitigates the adverse effects of inter-SNO handovers and facilitates seamless LEO constellation collaboration. The key idea behind Satitch is to decouple access satellites from individual SNOs’ core networks, enabling flexible association between shared satellite access points and multiple core networks. This design choice allows terrestrial users to remain connected to their original SNOs as much as possible. Concretely, Satitch incorporates a flexible, dynamic scheduling algorithm for space-ground connections that aims to minimize inter-SNO handovers experienced by terrestrial users. Further, Satitch adopts a proactive handover optimization to alleviate the impacts of inevitable interruptions. Our evaluation based on real-world constellation data demonstrates that Satitch can reduce up to 99.87% flow-level disruption time as compared to existing approaches and ensure smooth real-time video playback.

Wi-Fi sensing technology enables non-intrusive, continuous monitoring of user locations and activities, which supports diverse smart home applications. Since different sensing tasks exhibit contextual relationships, their integration can enhance individual module performance. However, integrating sensing tasks across different studies faces challenges due to the absence of: 1) a unified architecture that captures the fundamental nature shared across diverse sensing tasks, and 2) an extensible pipeline that accommodates future sensing methodologies. This paper presents UNI-FI, an extensible framework for multitask Wi-Fi sensing integration. This paper makes the following contributions: 1) we propose a unified theoretical framework that reveals fundamental differences between single-task and multi-task sensing; 2) we develop a scalable sensing pipeline that automatically generates a multi-task sensing solver, enabling seamless integration of multiple sensing models. Experimental results show that UNI-FI achieves robust performance across tasks, with a median localization error of approximately 0.54m, 98.34% accuracy for activity classification, and 98.57% accuracy for presence detection.

Acoustic communication is becoming increasingly pivotal in underwater environments, which are often challenged by strong multipath propagation and high ambient noise. However, existing solutions typically depend on specialized hardware, complex modulation schemes, or increased transmission power to ensure reliability, which are costly and unsustainable for energy-constrained underwater systems. To overcome these limitations, we propose DeWater, a deep learning-enhanced Chirp Spread Spectrum (CSS) demodulation framework designed for robust and energy-efficient communication. DeWater integrates a Swin Transformer-based demodulator with two novel components: a Frequency Transformation Block (FTB) and a Wavelet Convolution (WTConv) module, which together enable accurate demodulation without increasing transmitter complexity or power consumption. To further enhance noise resilience, we propose a two-stage fine-tuning strategy: Stage 1 applies noise-aware data augmentation through temporal and spectral masking, and Stage 2 employs adversarial training using Gradient Norm Adversarial Augmentation (GNAA). Extensive real-world underwater experiments using scenario-specific training demonstrate that DeWater achieves a significant performance gain, yielding an 82.81% reduction in symbol error rate compared to traditional dechirp demodulation. Notably, to maintain a 10% symbol error rate, DeWater requires 66.52% less transmitter power than the baseline, demonstrating its effectiveness as a low-power, robust solution for underwater acoustic communication.

We study how to update cache content to efficiently serve demands for dynamic data items that are frequently refreshed at backend servers. A significant challenge for optimizing this new type of cache systems with dynamic content is the following "information barrier". Because the precise data refresh events happening at the backend are not known to the cache, it is difficult for the cache to decide how to update cache content effectively. To overcome this information barrier, we introduce a new model that allows caches to observe data-item freshness at backend servers occasionally by paying an observation cost. A caching optimization problem is then formulated, aiming at strategically utilizing the observation option to balance the trade-off between providing fresh content and reducing data traffic to the backend. Although this optimization problem can be cast as a multi-action Restless Multi-Armed Bandit (RMAB) problem, it unfortunately does not satisfy the standard notion of multi-action indexability, preventing the use of standard Whittle index policies. We address this difficulty by introducing two new concepts of super-action indexability and sub-action indexability. These new notions of indexability significantly generalize the classical notion of multi-action indexability, and they enable us to develop low-complexity and asymptotically-optimal index-like policies for this otherwise intractable problem. Extensive numerical simulations verify that the proposed new policies benefit from more informed decision-making through strategic observation and significantly outperform existing benchmarks that do not exploit observation opportunities.

We present Pendulum, a live video analytics system with a novel network-compute joint scheduling in mobile edge computing (MEC) architecture. In practical scenarios, resource bottleneck frequently alternates across network (video streaming) and compute (DNN inference) stages due to independent fluctuations of wireless channel and scene content. However, prior single-stage scheduling systems suffer from throughput/accuracy fluctuation and resource wastage due to over-provisioning. To overcome the limitations, we newly leverage the interplay between video bitrate and DNN complexity to design an end-to-end system composed of (i) a resource-efficient network-compute demand curve profiler and (ii) a joint resource scheduler. Evaluation with various videos and state-of-the-art DNNs show that Pendulum achieves up to 0.64 mIoU gain and 1.29 × higher throughput than state-of-the-art baselines. Pendulum also achieves near-optimal multi-user resource scheduling performance with minimal search overhead, achieving 25% cost reduction compared to the network-compute decoupled scheduling.

The growing demand for deploying large language models (LLMs) in applications requiring low latency and strong privacy has brought increasing attention to edge inference. While distributing computation across edge devices helps alleviate resource bottlenecks, the sequential and layer-wise nature of autoregressive decoding still limits real-time deployment, leading to device underutilization, increased latency, and workload imbalance. Existing methods rely on heavyweight drafts and static policies, which cannot adapt to device heterogeneity and bandwidth variation, leading to limited parallelism and elevated latency. To bridge this gap, we propose EdgeSpec, a hierarchical speculative decoding framework at edges. Instead of relying on last-layer speculation, EdgeSpec enables intermediate transformer layers to emit draft tokens, which are validated and used to predict the next token in parallel across devices via attention masking. The system combines offline entropy profiling with an online Kalman filter to adaptively select confident trigger layers based on runtime load and input complexity. To further improve efficiency, EdgeSpec jointly optimizes model partitioning, entropy thresholds, and candidate sizes through a two-stage strategy that integrates offline planning with online adaptation. Experiments on LLaMA2 showed that EdgeSpec reduced inference latency by up to 17.1× while preserving output quality comparable to standard decoding.

In-network inference has emerged as a promising paradigm for enabling intelligent packet processing at the line rate within programmable data planes. However, it is fundamentally limited by an inherent conflict between model accuracy and the resource constraints of programmable data planes. This forces a compromise: monolithic deployments can achieve high accuracy but are constrained by the resource limits of a single switch, while distributed approaches leverage the combined resources of multiple switches but are limited in accuracy due to a lack of model coordination. We resolve this trade-off by proposing Monic, a framework that enables multiple "expert" submodels to perform collaborative inference. Inspired by the Mixture-of-Experts (MoE) paradigm, Monic uses a pipeline-compatible gating mechanism to selectively activate experts across the network. We enable this in practice through a resourceaware mapping and co-optimization strategy that automatically identifies optimal configurations under hardware constraints. We have implemented Monic using P4 hardware switches with Intel Tofino ASIC. Our evaluation shows that Monic achieves a 17.4% relative improvement over baseline methods and maintains a 29.58% Macro F1 advantage under scalability evaluation, demonstrating that a collaborative approach can simultaneously achieve resource efficiency and superior accuracy.

In-Network Aggregation (INA) has become an effective technique for accelerating distributed training by offloading gradient aggregation to programmable switches. However, existing INA protocols often suffer from the "Bucket Effect", where performance is limited by the scarcest resource along the transmission path, such as switch memory or link bandwidth. To overcome this limitation, we propose PAP, a progressive aggregation protocol that enables best-effort INA by dynamically coordinating switch memory and link bandwidth resources across the entire transmission path. PAP also addresses two critical reliability challenges: the mutual waiting lock problem caused by unresolved hash collisions, and incomplete aggregation due to non-deterministic packet processing. To this end, PAP incorporates a sequence-based mechanism for collision detection and a counter-based mechanism for identifying aggregation completion. Furthermore, it introduces a retransmission-aware recovery strategy to ensure correctness under packet loss. Experimental results on an FPGA-based testbed demonstrate that PAP can improve INA throughput by 21.2% to 132.2% and reduce distributed training time by 13.7% to 31.2% compared with state-of-the-art methods.

In the field of cybersecurity, it has been widely believed that air-gapped devices, disconnected from the Internet, offer strong protection against data breaches. However, recent research has challenged this assumption by demonstrating the feasibility of exfiltrating data from air-gapped devices using covert channels, such as modulating thermal, acoustic, or electromagnetic (EM) emanations. Existing approaches, however, face key limitations, including low bit rates, line-of-sight requirements, and susceptibility to EM shielding, limiting their practical deployments. To this end, we propose MagWhisper, a novel and practical data exfiltration framework that exploits GPU magnetic leakages for a covert channel. The non-radiative magnetic signals can penetrate various materials, including metal, enabling MagWhisper’s non-line-of-sight and shielded operation. We design a high-frequency frequency-shift keying modulation scheme over GPU magnetic signals, thus establishing a high-rate and robust channel. We further adopt a GPU-agnostic design to enhance MagWhisper’s generalizability. A proof-of-concept system is implemented to demonstrate MagWhisper’s feasibility. Our evaluation on three heterogeneous GPUs shows that MagWhisper achieves high bit rates exceeding 133 bps with a bit error rate (BER) lower than 1.2%. Furthermore, real-world case studies confirm that MagWhisper remains effective in non-line-of-sight scenarios and can even bypass EM shielding within a Faraday cage.

Low Earth Orbit (LEO) satellite constellations are becoming essential for expanding global Internet access, especially in remote and under-served areas. However, their highly dynamic nature, arising from network mobility, introduces complex coordination challenges between the dynamic satellites and the ground nodes (gateways and terrestrial devices). This is underscored by limited satellite visibility windows and spatially imbalanced user traffic demands. Local association (cell-satellitegateway) strategies, such as nearest-satellite or greedy load-based selection, result in partial terrestrial coverage or lead to load imbalance that affects traffic demand fulfillment. Network-driven orchestration through centralized optimization can strike an efficient balance between these two key objectives, but is often computationally intensive for periodic operation and real-time deployment. This work presents a learning-based network orchestration framework, NEO-GNN, that models a satellite-ground network as a dynamic spatio-temporal graph. In contrast to prior works, it employs a heterogeneous Graph Neural Network (GNN), where satellites, gateways, and ground cells are modeled as distinct node types to capture their varied visibility and networking capabilities. They are trained in an unsupervised manner using tailored loss functions to balance the dual requirements of coverage and utilization, and produce efficient, real-time association decisions during inference. Evaluations show that NEO-GNN delivers complete ground-cell coverage, improves traffic demand satisfaction through balanced satellite and gateway use, and remains robust under dynamic visibility and partial satellite failures. NEO-GNN provides a scalable and efficient alternative to traditional optimization methods for real-time network orchestration in bent-pipe LEO satellite systems.