INFOCOM 2026 Papers

We study the online sampling problem of a Wiener process, where the samples are transmitted to the remote estimator via an unreliable channel. The objective is to minimize the long-term average mean squared error (MSE) of the remote estimator subject to a sampling frequency constraint. To derive the online sampling strategy when the channel statistics is unknown, we reformulate the problem as an optimal stopping problem and review the sufficient condition of the offline MSE-optimal sampling threshold. Utilizing the sufficient condition, we then propose a stochastic approximation algorithm that adaptively approximate the optimum threshold. Our theoretical analysis shows the approximated threshold converges to the optimum sampling threshold almost surely. Moreover, let k be the number of successfully received samples, we proved that: the MSE gap between our online algorithm and the optimum offline algorithm decays with rate O(1/k), while for any online sampling algorithm that does not know the channel statistics in advance, the minimum MSE gap under the worst case delay distribution decays with rate more than Ω(1/k). To improve the speed of the adaptive learning algorithm, we integrate the Polyak–Ruppert (PR) averaging algorithm when updating the sampling threshold, achieving a faster convergence speed both theoretically and in simulations.

Traditionally, network devices detect DDoS attacks and network scans in real time by estimating the cardinality (or spread) of each destination or source IP from high-speed packet streams using per-flow cardinality estimators. However, modern attacks, such as carpet-bombing DDoS, distribute traffic across multiple IPs within a subnet, evading detection by reducing perIP spread. To address this challenge, we consider a new problem: identifying hierarchical super spreaders (HSSs) by estimating the spread of hierarchical flows in real time. Existing solutions either lack support for two-dimensional (2D) hierarchies or suffer from suboptimal accuracy. In this paper, we propose H-MOPS, a new sketch that accurately detects both one-dimensional (1D) and 2D HSSs. H-MOPS organizes hierarchical flows into a grid structure based on IP mask lengths, deploying a MOPS sketch at each node to estimate per-flow spread. MOPS improves accuracy by maintaining a prefilter to separate top-k super spreaders and supports mergeability across distributed sketches. To compute the conditional spread of a hierarchical flow, H-MOPS addresses the duplicated-counting problem by merging the virtual estimators of all its descendant HSSs and subtracting the merged result from its own spread. Experiments on CAIDA traces demonstrate that MOPS achieves lower estimation error and higher detection accuracy than existing methods, and that H-MOPS significantly outperforms prior approaches in both 1D and 2D HSS detection.

Checkpointing is the most prominent mechanism for large-scale deep learning training to ensure reliability, against frequent failures. Existing checkpointing systems suffer from significant communication overhead of transmitting checkpoints to remote storage or remote CPU memory. This leads to checkpointing-induced stalls, which then slows down the training. This paper presents Zoetic, a novel checkpointing system that performs checkpointing at every iteration without impairing training efficiency. The key idea behind Zoetic is to enable a machine to update remote checkpoints in its CPU memory by leveraging regular training traffic, inherent in gradient synchronization of data-parallelism. In contrast to traditional checkpoint transmission, Zoetic reduces checkpointing traffic to one-sixth or even zero, effectively eliminating checkpointing-induced stalls. We orchestrate the transmission of gradients from GPU HBM to CPU memory to avoid interfering with regular training traffic, and design an enhanced CPU checkpoint optimizer to overlap its operation with GPU training computations. We further propose a fast recovery method, which utilizes direct remote GPU access and non-serialized transmission to reduce recovery latency. Experimental results show that Zoetic achieves iteration-wise checkpointing without prolonging training time, while existing approaches slow down training by up to 41.5%. Additionally, Zoetic exhibits an 11.7 × faster checkpoint loading efficiency.

Caches serve as the cornerstone of latency-sensitive systems, yet a recently discovered phenomenon, so-called delayed hits, challenges conventional optimal caching algorithms, rendering them incapable of guaranteeing minimal latency. To fill this gap, in this paper, we pioneer a generic algorithmic analysis for distributed caching under delayed hits, making three contributions. Firstly, we analytically establish tight competitive ratio lower bounds for both deterministic and randomized online caching algorithms. Secondly, we develop a novel online caching optimization framework by generalizing the Least Recently Used (LRU) policy to distributed caching with delayed hits, encompassing both deterministic and randomized paradigms. Our key technicality is a timer-based data fetching scheme that leverages query history to limit aggregate data retrieval latency. We prove that our deterministic and randomized algorithms are both asymptotically optimal. Thirdly, we conduct extensive experiments on a real-world dataset to demonstrate the effectiveness and superiority of our algorithms over state-of-the-art solutions.

The proliferation of unauthorized autonomous aerial vehicles (AAVs) threatens public safety and privacy, especially elusive ones with tiny sizes, wide ranges, and unpredictable trajectories, making precise wide-range positioning challenging yet essential for threat mitigation. Existing approaches often rely on single-modal, single-view strategies, suffering severely degraded positioning error and identification accuracy from environmental adversity and limited observations. This paper presents MIROS, a distributed Multi-vIew Radar-vision cOgnitive fuSion framework architected around Observe-Orient-Decide-Act (OODA) cognitive loop for precise and robust elusive unauthorized AAV positioning. To close this loop, MIROS introduces two core innovations: (1) Orient: an abstract-binding-based fusion strategy that adaptively fuses heterogeneous sensor data via learned cross-modal and cross-view correlations; (2) Act: a situation-aware multi-view collaborative sensing strategy using dynamics modeling for proactive multi-view sensor coordination. Month-long deployments of MIROS over 3.16 × 108m3using minimal commodity sensors (one radar and three visual monitors) achieve 9.6 cm positioning error and improve identification accuracy by up to 2.14× over single-modal methods and 1.25× over single-view cross-modal baselines, while maintaining real-time inference and deployment scalability.

Homomorphic Encryption (HE) serves as a vital privacy-preserving technique in federated learning (FL) applications. However, while HE offers substantial privacy benefits, it also introduces significant computational and communication overhead. This challenge is further compounded in real-world heterogeneous scenarios, where clients vary in computing resources and privacy requirements. Existing solutions apply uniform static HE configuration to satisfy the privacy requirements of the most stringent client, resulting in high resource usage for all the clients. We propose AHE, a novel FL framework that integrates HE with adaptive privacy-aware clustering to dynamically optimize the latency-accuracy-privacy trade-off. AHE introduces two key innovations: (1) adaptive HE parameter management, which tailors encryption levels to individual client capabilities and customizable privacy needs, and (2) private cross-cluster aggregation, a lightweight key-switching based protocol that privately combines model updates across clusters with divergent HE schemes. To further minimize overhead, we design a parallelized key-switching mechanism that reduces server-side latency. Experimental results demonstrate that our approach enhances various FL applications by improving final accuracy up to 19% and reducing global round time up to 70%, significantly boosting training efficiency. Additionally, AHE is robust to dynamic client shift, and the parallelization strategy dramatically reduces server-side latency, enabling better scalability.

Ethereum 2.0 has adopted Proof-of-Stake (PoS) consensus to replace the energy-intensive Proof-of-Work (PoW) consensus in Ethereum 1.0. Under PoS, each block must be processed within a fixed 12-second slot, with a strict 4-second propagation window to ensure security and stability. Fast block propagation is therefore critical not only to meet this requirement but also to maximize throughput by allowing more transactions per block. However, our investigation on Ethereum 2.0 MainNet reveals that approximately 9% of blocks exceed the 4-second propagation threshold, leading to forks that undermine network reliability. This propagation delay is primarily caused by two factors: (1) serial transaction execution, which prolongs validation time; and (2) large transaction payloads, which increase transmission time. To address these challenges, we propose FBP-Eth2.0 (Fast Block Propagation in Ethereum 2.0), a novel approach that integrates two core innovations: (1) a practical parallel transaction execution scheme by classifying transactions into independent and dependent sets to reduce validation latency; and (2) a proactive compact block protocol that leverages the fixed 12-second slot to proactively exchange transaction pool snapshots among nodes, precisely identifying and replacing redundant transactions with ultra-short hashes. This proactive compaction design compresses the block while eliminating the additional communication overhead typically associated with compact block protocols, significantly improving transmission efficiency. Extensive experiments on Ethereum 2.0 MainNet demonstrate that FBP-Eth2.0 accelerates block propagation by up to 3× over existing schemes. Importantly, it reduces the ratio of blocks exceeding the 4-second threshold to below 1%.

Video surveillance plays a crucial role in modern society, fostering safer, smarter, and more efficient environments. However, it is of great challenge to transmit, store and analyze city-scale surveillance videos due to the massive amounts of data. In this work, we propose Juice, a lightweight surveillance video compression scheme that can be implemented on H.265-compliant cameras. The core idea of Juice is to effectively utilize the CU (Coding Unit) division information generated during the encoding process to predict tiles with foreground objects in each frame. Furthermore, redundant background tiles between frames are removed to minimize the compressed video size without compromising the downstream surveillance detection accuracy. We collect a real-world transportation traffic surveillance video datasets, consisting of 541 video clips recorded at 42 distinct locations in different light and weather conditions. We implement Juice and conduct extensive trace-driven simulations. The results demonstrate that Juice is lightweight and can process at least 32 FPS at a resolution of 1920×1080 on a single-core common CPU and can achieve an average 78.2% compression rate. Furthermore, Juice has little impact on downstream AI detection algorithms. Specifically, the object detection on compressed videos undergoes a minor 1% accuracy drop, compared to detection on uncompressed videos.

The constrained combinatorial multi-armed bandit model has been widely employed to solve problems in wireless networking and related areas, including the problem of wireless scheduling for throughput optimization under unknown channel conditions. Most work in this area uses an algorithm design strategy that combines a bandit learning algorithm with the virtual queue technique to track the throughput constraint violation. These algorithms seek to minimize the virtual queue length in their algorithm design. However, in networks where channel conditions change abruptly, the resulting constraints may become infeasible, leading to unbounded growth in virtual queue lengths. In this paper, we make the key observation that the dynamics of the head-of-line age, i.e. the age of the oldest packet in the virtual queue, make it more robust when used in algorithm design compared to the virtual queue length. We therefore design a learning-based scheduling policy that uses the head-of-line age in place of the virtual queue length. We show that our policy matches state-of-the-art performance under i.i.d. network conditions. Crucially, we also show that the system remains stable even under abrupt changes in channel conditions and can rapidly recover from periods of constraint infeasibility.

Collaboration-based personalized federated learning (Co-PFL) achieves superior accuracy on non-IID data by leveraging a collaboration graph that captures pairwise similarities among client data distributions for personalized model aggregation. However, its synchronous operation suffers from substantial delays caused by slow clients, limiting its applicability in latency-sensitive IoT applications. A natural alternative is asynchronous Co-PFL, where the server processes model updates as they arrive. Yet, asynchrony introduces staleness that distorts both collaboration graph estimation and personalized model aggregation, two tightly coupled components in Co-PFL. Existing asynchronous federated learning methods only handle staleness in global model aggregation and fail to maintain this coupling. To bridge this gap, we propose PACE, the first asynchronous Co-PFL framework that jointly manages staleness across collaboration estimation and model aggregation. PACE employs two lightweight, theoretically grounded server-side mechanisms: a collaboration-aware buffer update that adaptively refreshes buffered models based on staleness and collaboration relevance, and a staleness-triggered model multicast that selectively disseminates fresh aggregates when collaboration estimation error exceeds a threshold. Extensive evaluations show that PACE matches the accuracy of state-of-the-art synchronous Co-PFL methods, while reducing convergence time by up to 8.58×, enabling practical Co-PFL with IoT devices.

Encryption has been commonly used in network traffic to secure transmission, but it also brings challenges for malicious traffic detection, due to the invisibility of the packet payload. Graph-based methods are emerging as promising solutions by leveraging multi-host interactions to promote detection accuracy. But most of them face a critical problem: Graph Drift, where the flow statistics or topological information of a graph change over time.To overcome these drawbacks, we propose a graph-assisted encrypted traffic detection system, MalMoE, which applies Mixture of Experts (MoE) to select the best expert model for drift-aware classification. Particularly, we design 1-hop-GNN-like expert models that handle different graph drifts by analyzing graphs with different features. Then, the redesigned gate model conducts expert selection according to the actual drift. MalMoE is trained with a stable two-stage training strategy with data augmentation, which effectively guides the gate on how to perform routing. Experiments on open-source, synthetic, and real-world datasets show that MalMoE can perform precise and real-time detection.

VPN fingerprinting techniques are essential for network administrators to mitigate security threats arising from unauthorized or harmful VPN usage. Existing methods often rely on complex, protocol-specific signatures or computationally expensive AI models, limiting their practical applicability. In this paper, we introduce VPNSpotter, a practical VPN fingerprinting technique that infers coarse-grained field specifications from consistent data channel packets. VPNSpotter employs heuristic filtering to discard control channel packets and inconsistent packets caused by retransmission, segmentation, and aggregation. It then infers coarse-grained field specifications in a column-wise manner, categorizing fields into five predefined types. Evaluations on diverse VPN protocols—including mainstream, proprietary, and obfuscated variants—demonstrate that VPNSpotter accurately and efficiently identifies VPN traffic, outperforming prior signature-based and AI-based approaches.

Sparse sensing is a critical task in many real-world systems, heavily relying on effective data imputation. However, most existing methods overlook the severe sparsity of historical data available for training, which poses a fundamental challenge to achieving accurate imputation. Intrinsic priors-based models (i.e., MF, ODEs) are robust to sparsity but have limited performance with abundant historical data, whereas deep learning models excel with rich data but fail dramatically when data are sparse. Furthermore, this data sparsity introduces significant solution uncertainty and creates a need for more computationally efficient dependency modeling. To address these, we propose SH-Imputer, a novel adaptive framework. It first employs our proposed VDMF and AVDCDE modules to model the uncertainty of intrinsic spatiotemporal properties. Subsequently, a lightweight ST-Mamba module efficiently learns complex spatiotemporal dependencies. The entire process is governed by an adaptive mechanism that balances prior-based robustness with data-driven expressiveness, ensuring superior performance under varying historical data sparsity. We also provide theoretical justifications for our core design choices. Extensive experiments validate that SH-Imputer significantly outperforms state-of-the-art methods. Our code is available at https://github.com/JLUDhhh/SH-Imputer.

The rapid growth of mobile IoT is increasingly driven by eSIM technology, which enables seamless switching between network operators. Despite its potential, large-scale performance evaluations of commercial eSIM systems remain scarce. This study presents the first comprehensive analysis of eSIM reliability, utilizing a longitudinal dataset of over 22 million IoT devices in China over four years. Our analysis reveals that current failover-based multi-operator access (FAMOA) strategies significantly hamper performance due to high handover latency and disconnection risks. To address these limitations, we propose NeSIM, a framework implementing a novel Lookahead-based Multi-Operator Access (LAMOA) strategy. By leveraging AI-driven predictive analytics to proactively orchestrate operator switches, NeSIM reduces the risk of disconnectivity by 80.63% and cuts average handover latency from 9.93 to 5.26 minutes.

To support elastic services with flexible processing configurations in distributed computing networks, we propose a joint network resource and service control policy that optimizes routing, scheduling, processing locations, and the selection of processing configurations. The flexibility of service configurations introduces a new degree of freedom for improving network performance. First, we formulate the maximum elastic service flow problem and develop a polynomial-time algorithm to compute the network’s maximum throughput for supporting elastic services that can be processed by any available chain of service functions. Under stochastic service arrivals, we design a dynamic control policy that, upon each service arrival, chooses the service configuration, processing locations, and routing. In addition, we develop a scheduling policy to maximize network throughput. Furthermore, we extend the policy to maximize network utility by assigning a utility value to each service configuration. We prove that this utility maximization policy achieves an expected total network utility that is near-optimal.

Federated learning (FL) enables multiple clients to collaboratively train a machine learning model without sharing their local data. However, the distributed nature of FL makes it vulnerable to backdoor attacks from malicious clients. Most existing attack methods often assume that attackers can inject backdoors in every training round - a scenario that is both unrealistic and inefficient in real-world FL deployment. In this paper, we investigate why backdoor attacks become less effective under low-frequency injection and propose a novel attack paradigm for FL, called REinforced Memorization-based INterval backDoor attack (REMIND). REMIND optimizes the backdoor trigger via task alignment and feature alignment. Task alignment aligns backdoor and main task objectives to resist benign update suppression during non-attack rounds, while feature alignment guides poisoned samples to match the activation trajectory of target-class samples. This dual alignment enhances the backdoor's persistence and narrows the divergence between malicious and benign updates. With strong attack success rates established, we further analyze the advantages of low-frequency backdoor attacks, particularly their ability to improve robustness against defense mechanisms. Extensive evaluations on four benchmark datasets show that REMIND consistently outperforms eight state-of-the-art attack baselines under nine defense strategies.

Virtualized environments are the foundation of modern computing, but their complex and dynamic virtual networks make failure diagnosis challenging. First, failures are often triggered by specific network flows, but generic probes cannot replicate these flows to reproduce and diagnose the fault. Second, existing diagnostic tools lack the required granularity, operating at the device or IP level and thus missing faults that occur at the level of network device ports (not the port concept in TCP or UDP). To fill the gap, we propose and implement VNetPath, an approach that directly diagnoses failures by replaying the network context of a failed request and constructing its end-to-end paths, operating at the granularity of network device ports with eBPF-based non-intrusive tracing. We conducted experiments by selecting typical network fault types from a real-world cloud system, designing fault injection methods, and injecting 6942 faults to a testbed cloud. Our experimental results show that VNetPath achieves a diagnosis accuracy of 93.45% on injected faults, demonstrating its ability to pinpoint failures that are invisible to telemetry systems like Zoonet (73.27%) and Traceroute (18.22%), while incurring a low standing overhead of only 0.81%.

Compressed Sensing (CS), a sparse signal compression technique, has been adopted in federated learning (FL) reduce to network communication overhead. However, the computational cost incurred by CS reconstruction is commonly ignored, often surpassing the benefits derived from communication compression. This inefficiency arises from two main factors: (1) current CPU architectures are ill-suited for efficiently performing critical reconstruction operations, leading to low computational efficiency; and (2) reconstruction is executed centrally and sequentially at the server, imposing substantial computational burdens when facing multiple clients. To address these issues, we propose HiFedCS, a novel framework enabling accelerated and distributed CS reconstruction in FL. HiFedCS comprises two key components: (1) a hardware layer, optimize the critical operators in the reconstruction process at the circuit level; and (2) a software layer, offload the centralized reconstruction task from the server to clients for distributed execution. We implemented a HiFedCS prototype on a Xilinx Alveo U55C FPGA. Experimental results show that HiFedCS achieves up to 7.57× and 6.61× speedups over CPU and GPU, respectively, in gradient reconstruction tasks, and provides 2.52× and 1.92× end-to-end acceleration in real-world FL applications.

Multi-agent reinforcement learning (MARL) has been widely adopted in embodied artificial intelligence (embodied AI) systems, such as cooperative robotics, autonomous swarms, and wireless-enabled intelligent agents, where decision-making relies on local perception, physical interaction, and limited communication. However, most existing MARL algorithms require centralized training with global state-action information, which is often infeasible in real-world embodied settings due to communication unreliability and execution constraints. In this paper, we propose a distributed and communication-efficient MARL algorithm for embodied multi-agent systems. Each agent employs an approximated policy gradient using only its own action and locally available state and reward information. To further reduce communication and computation costs, we adopt a linear function approximation based on κ-hop neighbor states, naturally matching the local perception and interaction range of embodied agents. We prove global convergence with bounded errors induced by unreliable communication. Simulation results show that our method closely approaches centralized performance while achieving up to 99.3% runtime reduction, and outperforms Soft Actor-Critic (SAC) with a 76.0% reduction in computation time. The source code of the proposed method is available at: https://github.com/Pengcheng-Dai/DACA.