INFOCOM 2026 论文列表

Quantum key distribution (QKD) improves security in high-capacity space division multiplexing-based elastic optical networks (SDM-EONs) with multi-core fibers (MCFs). However, integrating QKD introduces inter-core crosstalk (IC-XT) and additional noise from quantum and classical channels, increasing blocking probability (BP) and complicating analytical evaluation. To address this, for the first time, we propose an exact analytical priority-based continuous-time Markov chain model, called QKD-Analytical, for QKD-enabled SDM-EONs. It computes BP by accounting for IC-XT and channel noise, while prioritizing cores based on node adjacency to optimize quantum channel placement and resource utilization. The model generates all feasible states and transitions to determine state probabilities accurately. It supports both single-class and multi-class traffic — using uniform slots for the former and variable slots for the latter to meet diverse bandwidth demands. To ensure scalability, we also develop an iterative approximation model for single-hop link analysis. Numerical results indicate that QKD-Analytical yields BP and resource utilization that closely match those obtained by Monte Carlo simulations, while the approximation model maintains good accuracy in larger scenarios. Additionally, QKD-Analytical outperforms non-priority-based models in terms of BP, highlighting the benefits of priority-aware core selection in QKD-enabled SDM-EONs.

The rapid proliferation of the IoT and AI has heightened the demand for non-intrusive, high-precision sensing in smart home environments. Acoustic-based methods offer lower costs and enhanced privacy but often struggle to achieve high accuracy, long-range, and wide-area functionality. This paper presents EchoGuard, an acoustic-based system for room-scale user tracking and gait recognition using commercial smart speakers. EchoGuard adopts an OFDM-based based demodulation framework that exploits phase and amplitude variations across subcarriers to enable accurate 2D localization, effectively mitigating truncation effects in conventional acoustic localization methods. To support real-time deployment, EchoGuard further incorporates dimensionality reduction and pruning techniques, reducing the computational overhead of localization and motion imaging. By modeling continuous human motion and suppressing variations in user position and orientation, EchoGuard enables spatial-state-independent motion imaging for robust gait recognition. Experimental results show median localization errors of 0.18m and 0.16m in static and dynamic scenarios, respectively, and achieve a gait recognition accuracy of 79.2% across 10 users, consistently outperforming existing baseline approaches.

Low Earth Orbit (LEO) satellite networks are emerging as a promising solution for providing global Internet access. However, their expansive scale and dynamic topologies present significant challenges for network management and routing optimization. While software-defined satellite networks (SDSNs) provide centralized control and global visibility, their effective operation hinges on accurate modeling of network states under constrained control-plane resources. Existing learning-based methods typically rely on resource-intensive link-level measurements, which are impractical in LEO constellations due to limited telemetry and high synchronization overhead. In this paper, we propose DS-Route, a graph neural network-based framework for flow-level delay prediction that relies solely on flow attributes and routing information, eliminating the need for fine-grained link-level measurements. To address the scalability and topological complexity of large-scale LEO satellite networks, DS-Route incorporates a dual-scale modeling strategy that integrates global topological embeddings with local path-level dependencies. Comprehensive evaluations using Kuiper and Starlink topologies demonstrate that DS-Route reduces mean absolute relative error (MARE) by 15.1% and 22.6%, respectively, compared to best-performing baselines, while maintaining robust performance under realistic satellite conditions. Furthermore, DS-Route’s routing performance is validated through consistent improvements in key Quality of Service indicators compared to existing predictors, highlighting its potential for effective satellite traffic engineering.

Next-generation wireless networks operating in mmWave and sub-THz bands promise to deliver unprecedented data rates in densely populated networks, enabling advanced applications such as immersive AR/VR and UAV communications. However, achieving this objective requires not only wide bandwidth, but also efficiency gains of multi-user (MU) multiplexing vs. serving one user at a time. While MU multiplexing is relatively well understood and widely deployed below 7 GHz, such solutions rely on the rich scattering environment of lower frequency channels. To date, beyond simple polarization multiplexing, there is no same-channel MU solution for line-of-sight channels that are often encountered at higher frequencies. In this paper, we introduce Multi-Spotlight, a novel system for sub-THz MU networking to generate multiple, finely concentrated spotlights, each precisely aligned with each user’s location. To do so, we exploit near-field wavefront engineering available due to small wavelengths relative to the aperture size and design high-resolution metasurfaces with phase profiles that combine linear and quadratic phase distributions, optimized to focus each user’s signal precisely at their location. We minimize inter-user interference by reducing overlap between the spatial footprints of the spotlights, thereby enabling the multiplexing of users even with the same angle of departure (AoD) at different distances from the transmitter. Our experimental results demonstrate Multi-Spotlight’s ability to multiplex users sharing the same AoD with inter-user separation as small as 50 wavelengths, and to deliver significant MU gains even under low SNR conditions, ensuring high aggregate network performance in challenging and dense environments.

The B+tree indexing scheme is widely employed in file systems and database systems. Modern data centers often adopt a disaggregated architecture, where compute servers and storage servers are deployed on separate machines connected via a network. Storage servers typically rely on B+tree-based indexing. However, with the rapid growth in network bandwidth, the bottleneck has shifted from the network to the indexing, challenging the scalability of these systems. In this paper, we present ForeIn, a novel architecture that offloads part of the storage indexing process to programmable switches within the network path. Specifically, we make three key contributions. First, we develop the PrefixCover algorithm, which converts a B+tree query into a longest prefix match query. This transformation allows partial deployment of B+tree operations in the switches at line rate. Second, we propose a greedy algorithm that dynamically adapts the PrefixCover algorithm to the resource constraints of programmable switches. Third, we design a data plane leveraging programmable switch capabilities, ensuring consistency between servers and switches with minimal overhead and minimal device modifications. We implement ForeIn on a testbed and conduct extensive experiments. Results demonstrate that ForeIn improves the throughput of B+tree-based storage servers by an average of 1.2 times. The source code is publicly available on GitHub [1].

We study Constrained Online Convex Optimization (COCO) in the two worlds of time-varying and time-invariant constraints. Our goal is to simultaneously minimize the regret to the best fixed solution in hindsight and the hard violation prohibiting any compensated constraint violations. We name the proposed COCO algorithm Double-Q, since it leverages both an original virtual Queue to construct a surrogate loss and a surrogate virtual Queue to construct a surrogate drift-plus-penalty (SDPP). This new Double-queue design enables tighter dual violation control compared to existing single-queue approaches. By minimizing an upper bound on the SDPP at each time, Double-Q achieves ${\mathcal{O}}(\sqrt T )$ regret and ${\mathcal{O}}\left({\min \left\{ {\sqrt T \log T,{T^v}} \right\}}\right)$ hard violation without requiring Slater’s condition, where v continuously quantifies constraint variation from 0 for fixed constraints to 1 for arbitrary constraints. For the first time, this hard violation bound simultaneously guarantees the best-known ${\mathcal{O}}(\sqrt T \log T)$ violation for arbitrary constraints at worst, and recovers the best-known ${\mathcal{O}}(1)$ violation for fixed constraints at best, thus bridging the best-known constraint violations in the two COCO worlds. For strongly convex loss functions, Double-Q achieves both improved ${\mathcal{O}}(\log T)$ regret and ${\mathcal{O}}\left({\min \left\{ {\sqrt {T\log T} ,{T^v}} \right\}}\right)$ hard violation. Simulation results demonstrate that Double-Q outperforms state-of-the-art COCO algorithms across various online applications.

Underwater Wireless Sensor Networks (UWSNs) generate critical data for marine exploration and monitoring, requiring efficient data collection mechanisms to ensure both data integrity and freshness. Autonomous Underwater Vehicles (AUVs), with their superior mobility and ability to establish close-range, high-bandwidth communication links with sensor nodes, have been widely adopted as mobile collectors. However, existing approaches rely on pre-planned trajectories that cannot adapt to dynamic factors. In this paper, we propose an intelligent scheduling framework that leverages Large Language Models (LLMs) to enable adaptive multi-AUV data collection in dynamic underwater environments. In particular, an iterative optimization algorithm is designed to refine solutions through multiple reasoning cycles, achieving efficient scheduling performance. Moreover, we propose a dynamic evaluation mechanism to enable prompt detection and adaptation to environmental changes, and introduce LoRA-based fine-tuning for autonomous performance evolution by leveraging historical task execution data. Comprehensive experiments demonstrate that our approach achieves at least 77.4% improvement in collected value of information (VoI) compared to state-of-the-art methods under dynamic scenarios.

PromptFL combines prompt tuning and federated learning to enable collaborative model training. It employs the entire Contrastive Language-Image Pre-training (CLIP) model on the client side, and submits the prompt learner for global aggregation. However, the large scale of the prompt network results in significant communication overhead per round, a challenge that remains unaddressed. To tackle this issue, this study presents a Feature Transmission-based PromptFL (FT-PromptFL), which distributes the CLIP encoder across clients and the server. Furthermore, we propose a Federated Feature Compaction (FFC) framework to minimize communication costs by constructing compact feature subsets through federated clustering. Theoretical analysis is provided to demonstrate the effectiveness of FFC. To mitigate performance degradation caused by compressed transmission, Semantic Compensation (SC) mechanism is designed through the multi-modal fusion of partial images and intact text. In addition, a dynamic weighting strategy is introduced to constrain the generalization error arising from asymmetric modality fusion. Thus, the model performance can be guaranteed while maintaining low communication cost. Experimental results show that the communication overhead of FT-PromptFL decreases 90% per-round compared to PromptFL. It also achieves superior inference performance, such as state-of-the-art (SOTA) 96.85% accuracy on the Flowers102 dataset.

Transmitting rich visual data in resource-constrained environments like Non-Terrestrial Networks (NTNs) poses a significant challenge. While current Semantic Communication (SC) approaches reduce bandwidth consumption, they often lack flexibility and/or compromise the perceptual fidelity of critical details. This paper first analyzes the fundamental trade-offs that exist between perceptual fidelity, semantic fidelity, and bandwidth utilization. It then introduces SPIFF, an SC-Generative AI framework that, by supporting selective fidelity, enables fine-grained control over the above trade-offs while meeting delay requirements. SPIFF features a lightweight, semantic-aware encoder performing semantic segmentation and applying a novel patch preservation strategy that retains perceptually significant regions while adapting lower-relevance areas compression to bandwidth availability. SPIFF also offloads high-complexity reconstruction tasks to a Generative AI-enabled decoder at the receiver, thus addressing asymmetric computation requirements. To support adaptation under dynamic conditions, while meeting system and application constraints, we equip SPIFF with a learning-based decision engine that is able to cope with the system non-linearities and effectively tune SPIFF’s configuration online. We evaluate SPIFF by implementing a full encoder-decoder pipeline. Results show that SPIFF fulfills perceptual reconstruction quality in scenarios where SC fails, and improves over state-of-the-art solutions both bandwidth savings (by up to 21%) and perceptual fidelity (by up to 13%).

Federated Learning (FL) remains highly vulnerable to poisoning attacks, especially under real-world hyper-heterogeneity, where clients differ significantly in data distributions, communication capabilities, and model architectures. Such heterogeneity not only undermines the effectiveness of aggregation strategies but also makes attacks more difficult to detect. Furthermore, high-dimensional models expand the attack surface. To address these challenges, we propose Horus, a heterogeneity-oblivious robust FL framework centered on low-rank adaptations (LoRAs). Rather than aggregating full model parameters, Horus inserts LoRAs into empirically stable layers and aggregates only LoRAs to reduce the attack uncover a key empirical observation that the input projection (LoRA-A) is markedly more stable than the output projection (LoRA-B) under heterogeneity and poisoning. Leveraging this, we design a Heterogeneity-Oblivious Poisoning Score using the features from LoRA-A to filter poisoned clients. For the remaining benign clients, we propose projection-aware aggregation mechanism to preserve collaborative signals while suppressing drifts, which reweights client updates by consistency with the global directions. Extensive experiments across diverse datasets, model architectures, and attacks demonstrate that Horus consistently outperforms state-of-the-art baselines in both robustness and accuracy.

The performance of federated learning (FL) over wireless networks critically depends on accurate and timely channel state information (CSI) across distributed devices. This requirement is tightly linked to how rapidly the channel gains vary, i.e., the coherence intervals. In practice, edge devices often exhibit unequal coherence times due to differences in mobility and scattering environments, leading to unequal demands for pilot signaling and channel estimation resources. Conventional FL schemes that overlook this coherence disparity can suffer from severe communication inefficiencies and training overhead. This paper proposes a coherence-aware, communication-efficient framework for joint channel training and model updating in practical wireless FL systems operating under heterogeneous fading dynamics. Focusing on downlink impairments, we introduce a resource-reuse strategy based on product superposition, enabling the parameter server to efficiently schedule both static and dynamic devices by embedding global model updates for static devices within pilot transmissions intended for mobile devices. We theoretically analyze the convergence behavior of the proposed scheme and quantify its gains in expected communication efficiency and training accuracy. Experiments demonstrate the effectiveness of the proposed framework under mobility-induced dynamics and offer useful insights for the practical deployment of FL over wireless channels.

In-band Network Telemetry (INT) has enabled efficient, real-time network monitoring and analysis by embedding telemetry data directly into live traffic. However, INT packet loss is inevitable due to network limitations (e.g., congestion, buffer overflow), which compromises the accuracy and effectiveness of INT-based applications. Existing INT recovery solutions cannot operate on the data plane, making them impractical for latency-sensitive in-network computing (INC) applications (e.g., RHA), which rely on accurate INT for decisions. To address this gap, this paper introduces SPRINT, the first forward-error-correction (FEC) INT recovery mechanism that runs entirely at line-rate within the programmable data plane. SPRINT leverages a novel, hardware-friendly shift powered XOR encoding with nested grouping to minimize redundant state and maximize throughput, thereby providing robust fault tolerance. We have implemented a testbed prototype of SPRINT on P4 hardware (Intel Tofino ASIC) and the BMv2 software switch. Our evaluation shows that SPRINT uses under 10% of on-chip SRAM, achieves a minimum recovery latency of approximately 21 µs, delivers a 99% recovery rate, and maintains strong reliability for INT-based applications, e.g., HPCC and RHA.

With over 3.5 million 5G base stations deployed globally, their collective energy consumption (projected to exceed 131 TWh annually) raises significant concerns over both operational costs and environmental impacts. In this paper, we present EExAPP, a deep reinforcement learning (DRL)-based xApp for 5G Open Radio Access Network (O-RAN) that jointly optimizes radio unit (RU) sleep scheduling and distributed unit (DU) resource slicing. EExAPP uses a dual-actor-dual-critic Proximal Policy Optimization (PPO) architecture, with dedicated actorcritic pairs targeting energy efficiency and quality-of-service (QoS) compliance. A transformer-based encoder enables scalable handling of variable user equipment (UE) populations by encoding all-UE observations into fixed-dimensional representations. To coordinate the two optimization objectives, a bipartite Graph Attention Network (GAT) is used to modulate actor updates based on both critic outputs, enabling adaptive tradeoffs between power savings and QoS. We have implemented EExAPP and deployed it on a real-world 5G O-RAN testbed with live traffic, commercial RU and smartphones. Extensive over-the-air experiments and ablation studies confirm that EExAPP significantly outperforms existing methods in reducing the energy consumption of RU while maintaining QoS.

Radio sensing in the sub-10 GHz spectrum offers unique advantages over traditional vision-based systems, including the ability to see through occlusions and preserve user privacy. However, the limited availability of spectrum in this range presents significant challenges for deploying large-scale radio sensing applications. In this paper, we introduce Ambient Radio Sensing (ARS), a novel Integrated Sensing and Communications (ISAC) approach that addresses spectrum scarcity by repurposing over-the-air radio signals from existing wireless systems (e.g., 5G and Wi-Fi) for sensing applications, without interfering with their primary communication functions. ARS operates as a standalone device that passively receives communication signals, amplifies them to illuminate surrounding objects, and captures the reflected signals using a self-mixing RF architecture to extract baseband features. This hardware innovation enables robust Doppler and angular feature extraction from ambient OFDM signals. To support downstream applications, we propose a cross-modal learning framework focusing on human activity recognition, featuring a streamlined training process that leverages an off-the-shelf vision model to supervise radio model training. We have developed a prototype of ARS and validated its effectiveness through extensive experiments using ambient 5G signals, demonstrating accurate human skeleton estimation and body mask segmentation applications.

Federated Learning (FL) enables collaborative model training across edge devices without sharing raw data. However, most FL approaches suffer from severe performance degradation when deployed to unseen edges with distribution shifts, known as the out-of-distribution (OOD) generalization problem. Given that causal relationships between inputs and labels are typically invariant under OOD scenarios, causality-inspired FL has attracted increasing attention. Existing works primarily focus on eliminating spurious correlations within each client’s local data while overlooking a critical step in FL, model aggregation. We show that in a structural causal graph, the aggregated model is a confounder leading to reliance on spurious correlations. To address this problem, we introduce a Federated Deconfounding learning (FedDEC) framework based on backdoor adjustment to eliminate the confounding paths. Theoretically, backdoor adjustment requires enumerating aggregated models under all possible aggregation weights, which is infeasible in practice. To efficiently approximate this process, FedDEC proposes a client-side disentangled feature learning module that separates personalized and global components. It also introduces a deconfounded training module that computes cross-attention between the personalized features and a server-maintained prototype bank to obtain deconfounded causal features. Experiments demonstrate that FedDEC achieves higher accuracy while maintaining comparable convergence speed compared to state-of-the-art algorithms.

Split federated learning (SFL) offers a promising solution for training large models on resource-constrained edge devices. However, existing synchronous and asynchronous SFL frameworks are plagued by a fundamental sequential execution pattern for communication and computation. This design creates a severe communication bottleneck, particularly for modern large models with substantial activation sizes, leading to significant resource underutilization and extended training time. To dismantle this bottleneck, we propose MicroSFL, a novel SFL framework with adaptive pipeline parallelism. MicroSFL enables workers to divide each of their training mini-batches into multiple small micro-batches and process them in a pipelined fashion. To counteract the challenges of system and statistical heterogeneity, MicroSFL adaptively assigns distinct micro-batch sizes to workers to accommodate their heterogeneous capacities, maximizing the temporal overlap between the computation and transmission of activations/gradients. Besides, to address statistical heterogeneity, MicroSFL employs an adaptive updating strategy. This strategy accumulates gradients at the server side to simulate updates on a large, balanced batch. Guided by a theoretical convergence analysis, the optimal updating weights are assigned to different workers to normalize their contributions for model updating. MicroSFL then dynamically determines the appropriate timing to apply the aggregated gradients, ensuring both stable convergence and high model accuracy. Extensive evaluations on a physical platform of 80 edge devices show that MicroSFL accelerates training by 3.64× to 5.67× and improves model accuracy by 3.6% to 6.1% compared to the baselines.

Function-as-a-Service (FaaS), the core implementation of Serverless Computing, is widely adopted in modern applications. However, its multi-tenant, shared infrastructure creates opportunities for network side-channel attacks, leading to the leakage of sensitive information such as function activity states. Most existing defense solutions lack specific considerations for the unique FaaS architecture. To address this issue, we propose FaaSGuard, an adaptive traffic obfuscation framework designed to protect function activity states and enhance the privacy of serverless applications. FaaSGuard achieves adaptiveness through a dual-mode monitoring mechanism, ensures covertness using traffic camouflage techniques, and guarantees application transparency with a tag-based packet handling mechanism. Our prototype evaluation on an open-source FaaS platform demonstrates that FaaSGuard effectively hides real traffic patterns under varying loads. The end-to-end response latency it introduces is almost negligible, and it achieves significant privacy enhancement at an acceptable resource cost. Specifically, it requires a maximum increase of about 25.3% in gateway CPU usage and a 12% increase in memory usage.

Urban spatio-temporal prediction is crucial for real-world applications such as traffic dispatching, environmental monitoring, and emergency response. Despite breakthroughs in deep learning that have advanced spatio-temporal pattern learning, the complex spatio-temporal dependencies inherent in urban dynamic systems continue to pose challenges for effective spatio-temporal modeling. Existing approaches often focus on modifying model architectures and assume that spatio-temporal information passes immediately, while neglecting the spatio-temporal delays caused by urban events, geographic factors, and sensor communication latency. To fill this gap, we propose CSTDFormer, which leverages contrastive learning to empower the pure Transformer architecture to capture both short-range and long-range spatio-temporal delays. Specifically, we introduce Spatio-Temporal Delay Contrastive Learning (STD-CL) combined with self-attention to model fine-grained short-range spatio-temporal delay effects, and then devise Local-Global Contrastive Learning (LG-CL) combined with cross-attention to model long-range spatio-temporal delay relationships. Extensive experiments demonstrate that our approach outperforms nine baseline methods on five real-world datasets from different urban spatio-temporal scenarios, highlighting the necessity of considering spatio-temporal delays in urban spatio-temporal prediction.

Active address probing is fundamental to large-scale network measurement, asset discovery, and security assessment. Although scenarios with seed addresses have witnessed substantial progress, research in seedless scenarios remains nascent, hindered by a lack of general patterns and the restricted applicability of probing methods, leading to a limited probing boundary and low probing efficiency. We propose 6Genos, a system for active IPv6 address discovery in seedless scenarios. Specifically, 6Genos begins by extracting patterns from seed addresses and identifying general ones to form a pattern library. It then constructs a General Pattern Genealogy Graph to model the hierarchical relationships among these patterns. Finally, it adopts a heuristic dynamic migration probing that adaptively incorporates feedback to achieve precise pattern matching and optimal resource utilization. Experimental results show that 6Genos achieves a hit rate of 1.32%~2.34%, outperforming state-of-the-art methods by 323% ~ 631%. Its coverage reaches 29.33%~31.83%, with 11%~46% improvement. Moreover, 6Genos completes large-scale probing 9.76× faster. By continuously probing all seedless BGP prefixes, 6Genos discovers 1.85B active IPv6 addresses across 46,156 BGP prefixes, significantly expanding the active address boundary and enabling broader-scale IPv6 research.