INFOCOM 2026 Papers

Next-generation wireless communications above 100 GHz promise significantly higher data rates, enabled by the wide swaths of available spectrum. To overcome the severe path loss in this regime, large-scale antenna arrays are required to close the link. As a byproduct of high carrier frequencies and large electrical sizes, the near-field region extends to meter-scale distances, allowing transmitters to focus energy directly onto receivers. However, despite this potential, conventional phased-array systems suffer from severe beam split, which causes frequency-dependent divergence of the beam and effectively limits the usable bandwidth at the receiver—even in line-of-sight scenarios. To address this challenge without resorting to costly hardware solutions, this paper introduces WiFocus, a novel phase manipulation technique that redistributes power more uniformly across frequency, thereby enhancing channel capacity. We show that a strategic adjustment of the beam focal point yields a closed-form, low-complexity, one-shot solution that achieves near-optimal performance. Extensive simulations and D-band over-the-air measurements confirm the effectiveness of the proposed technique, demonstrating average achievable channel capacity within 8.8% of the theoretical upper bound, a 50.0% improvement over conventional focused beams, and a 25.4% gain over the state-of-the-art phase manipulation method.

IP route lookup, a critical function in routers, switches, and cloud gateways, directly determining packet forwarding performance. However, IPv6 lookup faces significant performance challenges compared to IPv4 due to longer addresses and rapidly expanding Forwarding Information Bases (FIBs), resulting in slower speeds and higher memory consumption. Existing trie-and hash-based solutions fail to holistically accelerate lookup speed and reduce memory footprint.This paper introduces SkipTrie, a novel IPv6 route lookup scheme addressing these challenges through an innovative multi-bit trie search framework. SkipTrie employs sub-trie skipping that promotes indices of child nodes to their ancestors, eliminating intermediate node accesses during the tree traversal. This effectively increases search stride while reducing computation and memory accesses. Furthermore, to maximize stride with minimal per-node storage, SkipTrie features a node compression technique. This exploits the inherent sparsity of tries representing IPv6 FIBs, and leverages a dynamic programming algorithm to automatically select optimal node configurations. Extensive evaluation with FIBs from backbone and cloud networks demonstrates that SkipTrie is 1.5 ∼ 5.2× faster than state-of-the-art alternatives. In addition, it maintains excellent scalability across diverse FIB sizes, future FIB growth, and multi-core systems.

Time-sensitive networks, as in IEEE TSN, serve safety-critical applications and are found in closed networks, such as avionics or in-vehicle networks. Validating the timing requirements in such networks requires obtaining proven latency bounds by using, for example, the network-calculus framework. Although there exists a drive for scaling timesensitive networks to groups of networks with IETF DetNet, they do not scale beyond single-entity systems. Indeed, obtaining performance bounds requires the knowledge of all flows' and network's properties. This requirement is met when the system has a unique owner but is no longer satisfied when multiple stakeholders do not trust each other. In this paper, we propose a framework that enables network operators to provide latency guarantees for traversing flows using zero-knowledge proofs. These proofs attest to the honest execution of networkcalculus algorithms (total flow analysis and its variant TFA++) and guarantee the fulfillment of timing requirements without revealing sensitive data. Moreover, we develop a smart contract to manage contractual relationships, orchestrate payments, and verify the validity of submitted proofs. We also optimize the proof generation time and provide a proof of concept. Our scalability evaluations -including proof generation time and gas consumption -highlight the system's suitability for real-world deployment.

Mobile CrowdSensing (MCS) is a crowdsourcing-based paradigm that leverages mobile users to collect data from Points of Interest (PoIs) using their smart devices. As data freshness has become a crucial concern, Age of Information (AoI) is employed to measure data freshness for MCS systems. Existing AoI-aware MCS works mainly focus on direct data collection scenarios, where data in PoIs is always available. Unlike these works, this paper explores the AoI-aware MCS system for uncertain event capture applications, where events may occur frequently yet with uncertainty, making the AoI update hard to be estimated. First, we model this uncertain event capture problem as a constrained episodic restless bandit problem with unknown transition probability. Next, we propose a belief-DPP bandit policy by extending the Drift-Plus-Penalty (DPP) policy. By combining belief-DPP and the Thompson Sampling (TS) technique, we further propose the TS-DPP algorithm, so as to minimize the cumulative weighted AoI values of all events under a given budget constraint. We analyze the theoretical performance of the TS-DPP algorithm, and derive a sublinear Bayesian regret bound ${\mathcal{O}}\left( {\sqrt {T\log T} } \right)$, where T is the size of time horizon. Additionally, we conduct extensive simulations to demonstrate the significant performance of the TS-DPP algorithm.

This paper investigates two distinct types of block errors - undetected errors (confusions) and erasures - in additive white Gaussian noise (AWGN) channels with error-bounded block decoders operating in the finite blocklength (FBL) regime. While block error rate (BLER) is a common metric, it does not distinguish between confusions and erasures, which can have significantly different impacts in cross-layer protocol design, despite upper-layer protocols universally assuming physical (PHY) errors manifest as packet erasures rather than undetected corruptions - an assumption lacking rigorous PHY-layer validation. We present a systematic analysis of confusions and erasures under BLER-constrained maximum likelihood (ML) decoding. Through sphere-packing analysis, we provide analytical bounds for both block confusion and erasure probabilities, and derive the sensitivities of these bounds to blocklength and signal-to-noise ratio (SNR). To the best of our knowledge, this is the first study on this topic in the FBL regime. Our findings provide theoretical validation for the block erasure channel abstraction commonly assumed in medium access control (MAC) and network layer protocols, confirming that, for practical FBL codes, block confusions are negligible compared to block erasures, especially at large blocklengths and high SNR.

This paper addresses the challenge of deadline-aware online scheduling for jobs in hybrid cloud environments, where jobs may run on either cost-effective but unreliable spot instances or more expensive on-demand instances, under hard deadlines. We first establish a fundamental limit for existing (predominantly-) deterministic policies, proving a worst-case competitive ratio of Ω(K), where K is the cost ratio between on-demand and spot instances. We then present a novel randomized scheduling algorithm, ROSS, that achieves a provably optimal competitive ratio of $\sqrt K $ under reasonable deadlines, significantly improving upon existing approaches. Extensive evaluations on real-world trace data from Azure and AWS demonstrate that ROSS effectively balances cost optimization and deadline guarantees, consistently outperforming the state-of-the-art by up to 30% in cost savings, across diverse spot market conditions.

The escalating volume of network traffic drives the offloading of stateful Network Functions (NFs) to Data Processing Units (DPUs) for enhanced performance and host CPU relief. However, current DPU offloading approaches, while often boosting forwarding throughput, frequently suffer from poor concurrency performance. Effectively addressing this requires overcoming challenges inherent to DPU architectures, such as the Arm subsystem’s limitations in rapid connection setup and the intricate management of hardware flow table resources, which can otherwise cap connection processing rates (e.g., below 100k connections per second (CPS) with conventional methods). To address these limitations, we present ConnSched, a high-performance software framework for efficient and robust offloading of connection processing in stateful NFs to DPUs. ConnSched incorporates two key mechanisms: ConnSched-Filter adaptively identifies and filters out short-lived connections by monitoring their lifecycle and current DPU load, thereby reserving hardware resources for long-lived connections and enhancing robustness against concurrent traffic bursts. Complementing this, ConnSched-QoS implements a prioritized offloading mechanism, managing DPU hardware flow table entries by giving precedence to critical flows, especially when resources are constrained. Experimental results demonstrate that ConnSched achieves a 6.4× higher new connection creation rate compared to standard DPU offloading methods and 38.7% higher throughput than CPU-based approaches. ConnSched simultaneously improves concurrent connection handling capability while maintaining high packet forwarding performance.

Deploying large language model (LLM) on edge device enables personalized LLM agents for various users. The growing availability of diverse personalized agents presents a unique opportunity for peer-to-peer (P2P) collaboration, wherein each user can delegate tasks beyond the local agent’s expertise to remote agents more suited for the specific query. This paper introduces PPAI, the first personalized LLM agent interoperability system, which enables users to collaborate with each other based on agent specialization. However, the ever-changing pool of agents and their interchangeable capacity introduce new challenges when it comes to matching queries to agents and balancing loads, compared with existing P2P systems. Therefore, we propose a scalable query-agent pair scoring mechanism based on prototypes to identify suitable agents within a P2P network with churn. Moreover, we propose a multi-agent interoperability Bayesian game to balance local demand and global efficiency, when changes in remote agent load occur too quickly to be observed. Finally, we implement a prototype of PPAI and demonstrate that it substantially broadens the range of tasks that could be carried out while maintaining load balance. On average, it achieves an accuracy improvement of up to 7.96% across multiple tasks, while reducing latency by 16.34% compared to the baseline.

Autonomous exploration aims to navigate and map complex, hard-to-reach environments. To achieve effective co-ordination, robots rely heavily on reliable communication for real-time data sharing, collaborative decision-making, and synchronized movement. However, such missions are often hindered by significant communication challenges. Factors such as signal attenuation, fluctuating link quality, and intermittent connectivity often arise due to occlusion, increased distances, and radio interference. These uncertainties can lead to delayed or incomplete data transmission, ultimately reducing the exploration efficiency.This paper presents Nexus (Networked Elastic eXploration via Unified Swarm), a self-organizing robot swarm specifically designed to address communication uncertainties. As a decentralized and self-sufficient swarm framework, Nexus introduces two key innovations. First, Nexus enables proactive role differentiation, allowing robots to adaptively switch between "explorer" and "messenger" based on factors such as spatial distribution, predicted communication quality, and swarm topology. Second, Nexus pioneers the use of signal traces and local environmental sensing to predict communication channel conditions in a robot’s vicinity. This forward-looking capability guides strategic role assignments and task allocation. Integrating communication-aware decisions into path planning, Nexus enhances exploration efficiency by over 19.5% and significantly reduces instances of robot disconnection compared with SOTA methods, as demonstrated through extensive real-world and simulated experiments.

Randomized smoothing (RS) is a widely adopted certified defense offering provable robustness against adversarial attacks. Its plug-and-play nature, which requires only black-box access to the model, makes it well suited for protecting compressed and compiled models deployed on on-device platforms. However, its computational expense due to extensive Monte Carlo (MC) sampling limits practical deployment on resource-constrained platforms. We address this limitation by introducing the Agresti–Coull Style Studentized Bootstrap (ACS-Boot) interval, a computationally efficient alternative to the conventional Clopper–Pearson interval used in RS. ACS-Boot significantly reduces the sample complexity required to achieve comparable robustness guarantees, enabling certification at substantially lower cost. Experiments on MNIST, CIFAR-10, and ImageNet demonstrate that ACS-Boot maintains robust certification accuracy while cutting sample numbers by up to 20 times, achieving inference speed-ups of 20-24 times. Notably, under equal sampling budgets, ACS-Boot certifies up to 34% larger ℓ2perturbations, especially in low-sample regimes. We also provide theoretical insights into how the certified radius scales with sample size. Moreover, we validate ACS-Boot on edge and mobile devices, including Raspberry Pi and Google Pixel smartphones, achieving comparable certification accuracy with 10 times fewer samples and up to 10-fold faster inference, highlighting its real-world deployability.

Recent advances in underwater communication techniques have shown significant potential for supporting diverse applications. However, little work has been done to achieve reliable and ultra-long-range underwater communication using commodity mobile devices in various underwater environments. In this paper, we present HydroChirp, a novel acoustic-signal-based underwater communication system designed to achieve reliable and ultra-long-range communication using commodity mobile devices. To do this, HydroChirp dynamically selects the channels with minimal distortion and leverages a novel chirp segmentation technique to mitigate the impact of generic and low-bandwidth noises. To address frequency offsets, HydroChirp uses an innovative frequency-noise model to estimate and compensate for noise generated by water fluctuations and determine the optimal transmission power. Furthermore, HydroChirp enhances communication reliability and range by analyzing noise patterns generated by specific underwater objects. We implement HydroChirp on commodity mobile devices and extensively evaluate its performance across various underwater scenarios. The experimental results demonstrate that HydroChirp can achieve a communication range of up to 150 meters with high reliability and efficiency.

With AI workloads relying heavily on massive distributed datasets, the performance of remote file access has become crucial. Traditional remote file access mechanisms introduce significant overhead when retrieving data from storage, limiting performance in high-throughput applications. NVMe over Fabrics (NVMe-oF) is a remote file access mechanism designed for high-speed, low-latency access to remote block storage, making it a superior alternative to traditional solutions. However, NVMe is inherently a block-based protocol, meaning it does not natively support file names, directory structures, or POSIX semantics. In this paper, we design an approach to enable direct file-path reads over NVMe-oF, bridging the gap between high-performance block devices and file-oriented applications. Our approach streamlines data retrieval by leveraging NVMe’s high-speed architecture and low-latency capabilities, reducing unnecessary kernel interactions and optimizing I/O operations. Benchmark results show up to 2-14X improvement in remote read latency compared to existing approaches.

With the rapid progress in artificial intelligence (AI), the network edge has become a dominant source of context-rich data, which demands immediate processing and seamless sharing to optimize AI-as-a-service models. Edge AI (or edge intelligence) enables on-device inference and localized training without central coordination. This paradigm requires the timely exchange of model updates among distributed devices to ensure efficient collaboration and consistent model convergence under constrained resources. However, existing approaches are limited by bandwidth asymmetry, heterogeneous device churn, and model integrity under adversarial updates. Herein, we propose DcGossip, a data-centric gossip-based model update sharing approach that leverages information-centric networking to enable receiver-driven model retrieval with named, content-addressable updates, adaptive model discovery, and version-controlled exchange for communication-efficient collaborative edge learning. DcGossip also introduces a self-verifying model integrity verification mechanism to ensure secure propagation and tamper-proof validation of shared updates. Our extensive evaluations demonstrate that DcGossip achieves 41% faster convergence speed and 26% lower propagation latency, effectively reducing communication cost by 80% compared to existing approaches, while maintaining minimal verification overhead.

Wi-Fi tracking technology demonstrates promising potential for future smart home and intelligent family care. Currently, accurate Wi-Fi tracking methods rely primarily on fine-grained velocity features. However, such velocity-based approaches suffer from the problem of accumulative errors, making it challenging to stably track users’ trajectories over a long period of time. This paper presents DuTrack, a fusion-based tracking system for stable human tracking. The fundamental idea is to leverage the ubiquitous acoustic signals in households to rectify the accumulative Wi-Fi tracking error. Theoretically, Wi-Fi sensing in line-of-sight (LoS) and non-line-of-sight (NLoS) scenarios can be modeled as elliptical Fresnel zones and hyperbolic zones, respectively. By designing acoustic sensing signals, we are able to model the acoustic sensing zones as a series of hyperbolic clusters. We reveal how to fuse the fields of electromagnetic waves and mechanical waves, and establish the optimization equation. Next, we design a data-driven architecture to solve the aforementioned optimization equation. Experimental results show that the proposed multimodal tracking scheme exhibits superior performance. We achieve a 89.37% reduction in median tracking error compared to model-based methods and a 65.02% reduction compared to data-driven methods.

The Goal-oriented Communication (GoC) paradigm breaks the separation between communication and the content of the data, tailoring communication decisions to the specific needs of the receiver and targeting application performance. While recent studies show impressive encoding performance in point-to-point scenarios, the multi-node distributed scenario is still almost unexplored. Moreover, the few studies to investigate this consider a centralized collision-free approach, where a central scheduler decides the transmission order of the nodes. In this work, we address the Goal-oriented Medium Access (GoMA) problem, in which multiple intelligent agents must coordinate to share a wireless channel and avoid mutual interference. We propose a theoretical framework for the analysis and optimization of distributed GoMA, serving as a first step towards its complete characterization. We prove that the problem is non-convex and may admit multiple Nash Equilibrium (NE) solutions. We provide a characterization of each node’s best response to others’ strategies and propose an optimization approach that provably reaches one such NE, outperforming centralized approaches by up to 100% while also reducing energy consumption. We also design a distributed learning algorithm that operates with limited feedback and no prior knowledge.

Mobile edge computing (MEC) is pivotal for delivering low-latency services to IoT applications. However, inherent uncertainties in real-world applications, notably conditional execution workflows and variable service durations, are often overlooked, hindering efficient resource utilization and optimization flexibility. This paper introduces UTOC, an uncertainty-aware framework that collaboratively optimizes the execution of conditional directed acyclic graph (CDAG) applications in MEC networks by co-designing service deployment and task scheduling. First, we model service duration variability with the G/G/c queueing model and represent conditional workflows with probabilistic branching. Then we employ moment matching theory to approximate application end-to-end (E2E) latency, enabling the accurate quantification of the impact on E2E latency by service instances. This quantification enables the CDAG application execution optimisation problem to be decomposed into two subproblems, service deployment and task scheduling. Finally, we design a heuristic service deployment algorithm for efficient resource allocation and an online scheduling algorithm to minimize E2E latency while balancing server congestion within the MEC networks. Extensive experiments validate that UTOC effectively utilizes edge resources and significantly reduces E2E latency compared to state-of-the-art baselines.

With the rise of large language models, service providers offer language models as a service, enabling users to fine-tune customized models via uploaded private datasets. However, this raises concerns about sensitive data leakage. Prior methods, relying on differential privacy within device-cloud collaboration frameworks, struggle to balance privacy and utility, exposing users to inference attacks or degrading fine-tuning performance. To address this, we propose PrivTune, an efficient and privacy-preserving fine-tuning framework via Split Learning (SL). The key idea of PrivTune is to inject crafted noise into token representations from the SL bottom model, making each token resemble the n-hop indirect neighbors. PrivTune formulates this as an optimization problem to compute the optimal noise vector, aligning with defense-utility goals. On this basis, it then adjusts the parameters (i.e., mean) of the dχ-Privacy noise distribution to align with the optimization direction and scales the noise according to token importance to minimize distortion. Experiments on five datasets (covering both classification and generation tasks) against three embedding inversion and three attribute inference attacks show that, using RoBERTa on the Stanford Sentiment Treebank dataset, PrivTune reduces the attack success rate to 10% with only a 3.33% drop in utility performance, outperforming state-of-the-art baselines.

Virtual Network Embedding (VNE) is a fundamental yet challenging task in network virtualization, particularly in dynamic and uncertain environments. While deep reinforcement learning (DRL) has shown promise, its reliance on implicit knowledge modeling—entangling historical experience within neural network parameters—constrains adaptability and interpretability, hindering deployment in production networks. To address this, we introduce Knowledge-Driven VNE (KD-VNE), a framework built on the principle of decoupling knowledge representation from the policy network. At its core, KD-VNE leverages a Temporal Knowledge Graph (TKG) as an explicit world model, empowering the agent with structured reasoning over network dynamics. This externalized knowledge is integrated within a hybrid Actor-Critic agent that synthesizes real-time state, TKG-curated experience, and domain heuristics through a dual-loop mechanism, enabling both rapid tactical adaptation and long-term strategic optimization. Extensive experiments show that KD-VNE achieves competitive acceptance ratios and improved economic efficiency. More critically, it demonstrates exceptional cross-domain generalization, achieving the highest acceptance ratio on real-world network topologies—outperforming the DRL baseline by over 80%—all without retraining.

Network traffic classification, aiming to classify network traffic into distinct management groups, plays an indispensable role in network monitoring, QoS, and intrusion detection. The widespread of encryption techniques applied to networks has made accurate traffic classification increasingly challenging. In recent years, raw traffic data as input to deep learning models has been widely adopted due to its high classification accuracy and low reliance on human-engineered features. Yet, a critical challenge remains: different bytes within a packet carry varying levels of discriminative information, and no consensus exists on the optimal representation unit for model input.In this paper, deep analysis of encrypted network traffic data is conducted, and our studies show that as the input to deep learning models, (1) byte-level representation outperforms bit-level encoding; (2) unigram byte features surpass the other ’n’-gram variants (halfgram, bigram, and trigram); (3) packet header bytes are significantly more discriminative than payload bytes, especially in highly encrypted datasets; and (4) certain header fields exhibit higher informativeness. Based on these findings, we propose HF-Transformer, a lightweight, non-pretrained model that leverages optimal header fields for classification. Evaluated across six datasets, HF-Transformer achieves up to 6.04% higher F1-score than SOTA alternatives, demonstrating its efficacy for encrypted traffic classification.