INFOCOM 2026 论文列表

Permissioned blockchain systems provide a mutual-trust platform for data sharing and collaboration among organizations. However, performance bottlenecks limit their adoption in industrial applications requiring high transaction throughput and low latency. Recent advancements have focused on leveraging parallelism to improve performance, but transaction contention remains a significant challenge. In this paper, we introduce Overpass Ledger (OPL), a high-performance permissioned blockchain system that utilizes a highly parallelized workflow. To tackle transaction contention, we propose ReX, an advanced Optimistic Concurrency Control (OCC) mechanism that efficiently re-executes conflicting transactions, eliminating transaction abortion and maximizing resource utilization. By integrating ReX, OPL fully harnesses the advantages of parallel stage processing and concurrent transaction execution. Experimental results demonstrate that OPL achieves throughput improvements of 77×, 19×, and 4× compared to Hyperledger Fabric, BIDL, and FISCO BCOS, respectively, while maintaining consistently low latency.

Quantum networks are a promising technology for supporting a wide range of quantum applications with diverse service requirements, such as high fidelity, low end-to-end (E2E) latency, and high entanglement rate. However, in practice, the fidelity of entanglement is often unknown. Existing studies primarily focus on estimating fidelity and selecting the highest-fidelity path, but often overlook other application-specific requirements. To address this limitation, we propose Variance-aware Fidelity Threshold Path Selection (VaFTPS), an efficient candidate path selection algorithm that identifies paths satisfying fidelity requirements. VaFTPS exploits a key property of fidelity estimation: higher-fidelity paths yield lower-variance benchmarking measurements. It adopts a mean-variance reward formulation within a multi-armed bandit (MAB) framework. Based on benchmarking results, VaFTPS estimates entanglement fidelity of paths and eliminates paths unlikely to satisfy the fidelity threshold using a tailored confidence interval, thereby reducing resource consumption. To further support differentiated service, we design a request scheduling strategy that allocates requests to the selected candidate paths to maximize the total service rate for multiple requests while satisfying the fidelity, E2E latency, and entanglement rate requirements. Extensive experiments in NetSquid show that our approach efficiently identifies paths satisfying the fidelity requirement and provides differentiated service for various requests.

Crucially, pipeline parallelism (PP) is an indispensable parallelization strategy for training large models on multiple GPUs. Despite some proposed PP schedules theoretically promising zero bubbles, we highlight that runtime bubbles remain a significant practical hurdle in real-world model training. This discrepancy arises because these idealized schedules presuppose that F (forward), B (backward), and W (weight) operations have uniform and stable execution times. This assumption does not hold in practical scenarios.In this paper, we introduce AdaptPipe, an adaptive pipeline scheduling framework that mitigates runtime bubbles by leveraging our analysis of the characteristics of two types of runtime bubbles. AdaptPipe improves runtime efficiency by opportunistically filling W tasks of appropriate granularity. To achieve this, we develop a prediction method and a notification method to estimate the size of different runtime bubbles. Furthermore, we design a forward-backward (FB) scheme, namely nF –nB, which aims to provide AdaptPipe with sufficient schedulable W tasks to handle continuous or large bubbles while still adhering to memory constraints.Our experiments on dense models and sparse MoE models show that AdaptPipe improves throughput by 0.6%-13.8% over state-of-the-art static pipeline schedules while maintaining comparable activation memory usage.

Accurately measuring time passing is critical for many applications. However, in Trusted Execution Environments (TEEs) such as Intel SGX, the time source is outside the Trusted Computing Base: a malicious host can manipulate the TEE's notion of time, jumping in time or affecting perceived time speed. Previous work (Triad) proposes protocols for TEEs to maintain a trustworthy time source by building a cluster of TEEs that collaborate with each other and with a remote Time Authority to maintain a continuous notion of passing time. However, such approaches still allow an attacker to control the operating system and arbitrarily manipulate their own TEE's perceived clock speed. An attacker can even propagate faster passage of time to honest machines participating in Triad's trusted time protocol, causing them to skip to timestamps arbitrarily far in the future. We propose TriHaRd, a TEE trusted time protocol achieving high resilience against clock speed and offset manipulations, notably through Byzantine-resilient clock updates and consistency checks. We empirically show that TriHaRd mitigates known attacks against Triad.

Vision Transformers (ViT) achieve remarkable performance across a wide range of visual tasks but incur high inference latency on resource-constrained mobile devices. Mobile-cloud collaborative inference, i.e., partially offloading the workload to the cloud, is a promising solution for acceleration by leveraging both mobile and cloud resources. Existing approaches reduce token numbers in a multi-round manner to shrink the data to be offloaded to the cloud, so as to reduce the communication latency. However, they cause imbalanced workloads, leading to heavy latency on the resource-constrained mobile devices. This paper presents Mercury, a collaborative ViT inference framework that accounts for mobile-cloud resource disparity. Unlike multi-round strategies, it performs single-round token pruning at the initial layer on mobile devices to aggressively reduce computation and communication overhead, and then restores tokens in the cloud to maintain accuracy. A similarity-based regionally decentralized token pruning method is designed to preserve semantic features, while an efficient meta-parameter-based token reconstruction method is used to improve accuracy with minimal overhead. Experiments on real-world testbeds under 4G/5G traces show that Mercury achieves up to a 2.27× speedup and halves the mobile-side energy consumption, while preserving comparable accuracy to state-of-the-art methods.

Social network de-anonymization refers to the process where an adversary recovers user identities by mapping an anonymized network to a correlated real-name network, and has attracted research interests due to the growing privacy concerns. Prior art usually assumes that the adversary observes a single pair of correlated networks. However, real-world social networks naturally evolve, thus enabling the adversary to obtain correlated observations across snapshots. Although recent studies highlight the increased privacy risk in this temporal context, quantitative theory on how multi-snapshot observations affect the feasibility of de-anonymization remains unexplored.To address this, we investigate joint de-anonymization of temporal social networks. Given a social network modeled as a general random graph evolving over time, our goal is to map the anonymized users observed across multiple snapshots to their real identities. Specifically, we aim to answer two fundamental questions: (i) When is perfect joint de-anonymization possible? (ii) How to efficiently find the true mapping? To this end, we first theoretically establish both necessary and sufficient conditions for perfect joint de-anonymization, revealing a critical insight: leveraging multi-snapshot observations enables accurate user re-identification even if individual snapshots alone are insufficient. Then, by casting joint de-anonymization as a generalized quadratic assignment problem, we design efficient polynomial-time algorithms to approximate the true mapping based on convex relaxation and Frank-Wolfe optimization. Finally, extensive experiments on synthetic and real social networks validate our theoretical findings and the proposed algorithms.

Cross-Technology Communication (CTC) enables direct data exchange between heterogeneous wireless devices, but existing solutions lack security considerations, exposing billions of IoT devices to spoofing, replay, and jamming attacks. We present BeeKeeper, the first comprehensive security framework that protects bidirectional CTC between WiFi and ZigBee devices. BeeKeeper introduces a novel dual-binding mechanism that synergistically combines physical-layer channel reciprocity with lightweight cryptographic protocols. By extracting Instantaneous Frequency-Power Images (IFPI) from natural channel observations, our system provides location-specific authentication without pre-shared secrets. Moreover, the three-phase WiZBeeAuth protocol achieves mutual authentication, while frequency hopping and mesh coding ensure jamming resilience. Finally, extensive evaluation on commodity hardware demonstrates BeeKeeper’s effectiveness: 94.7% attack detection rate, 6.8% throughput reduction and 12.3% CPU overhead. Our work transforms vulnerable CTC into a secure foundation for mission-critical IoT applications, providing a practical solution for securing the heterogeneous wireless ecosystem.

Existing federated learning (FL)-based anomaly detection (AD) methods suffer from performance degradation due to domain heterogeneity, e.g., differences in features, data distributions, and data types across clients from different domains. This challenge is particularly pronounced in IoT systems and is more challenging than the well-known problem of non-i.i.d. data issues. To address this problem, we propose Personalized Federated Heterogeneous Autoencoder (PFAE), an AD framework that introduces a private head for processing domain-specific inputs and a public backbone for shared representation learning. PFAE design enables each client to effectively detect anomalies within its own data domain while leveraging global knowledge through model aggregation. We theoretically prove that the difference in the expected anomaly score of PFAE calculated using the public backbone for benign samples from any pair of clients is bounded. This implies that the distribution of benign samples is shared across domains, so benign samples from a client may not be misidentified as anomalies by other clients during inference. PFAE has lower training complexity than existing autoencoder-based methods. Experiments on eight non-i.i.d. datasets from different domains, where each client is trained on a single dataset, show that PFAE enhances the generalization of anomaly scores for benign samples across domains, especially under high anomaly ratios.

Extending RDMA across datacenters is hindered by inefficient loss recovery. High round-trip time (RTT) and large bandwidth-delay product (BDP) cripple the default Go-back-N (GBN) retransmission, while hardware constraints on RDMA NICs (RNICs) limit the feasibility of practical deployment of selective retransmission (SR). In this paper, we present LR2, an in-network collaborative retransmission mechanism that overcomes RNIC limitations to enable reliable long-distance RDMA. At its core, LR2 decouples retransmission responsibilities, assigning fast-loop GBN at the RNIC and reconstructing long-haul SR at datacenter interconnection (DCI) switches. In addition, DCI switches independently mask intra-DC losses through near-sender fast feedback and near-receiver local retransmission. To realize this design, LR2 deploys lightweight programmable modules at both ends of the long-haul link: (1) a near-receiver Depot module that maintains buffer-efficient reordering and backup pools, and (2) a near-sender Sentry module that generates fast loss feedback and filters out redundant retransmission. We implement LR2 on Tofino switches and evaluate it through testbed experiments and large-scale simulations. LR2 significantly reduces flow completion time (FCT) and improves bandwidth efficiency over baselines, without large system overhead.

End-to-end latency is a key performance metric for long-running dataflow systems, as it directly impacts the timeliness and quality of results in real-time applications. In cloud environments, dynamic resource scaling is essential for handling workload fluctuations efficiently. However, maintaining consistent latency remains challenging due to resource contention and scaling-induced disruptions. Existing scaling techniques improve efficiency but fall short in guaranteeing latency, as they lack accurate latency estimation and timely scaling decisions. We present Sluice, a general framework for guaranteeing end-to-end latency. The core insight behind Sluice is to decompose latency into intrinsic and extrinsic components, isolating predictable, controllable delays from external disruptions like scaling and scheduling. Sluice dynamically reserves a buffer for extrinsic latency and scales resources to tightly control intrinsic latency. At its core, Sluice integrates (1) a novel Latency Estimation Model (LEM) that uses instantaneous backlog sizes along the critical path, and (2) a scaling controller that leverages LEM to make latency-aware, resource-efficient scaling decisions. We implement Sluice on Apache Flink and evaluate it using diverse real-world workloads. Results show that Sluice consistently enforces end-to-end latency guarantees under dynamic conditions while maintaining competitive resource efficiency, outperforming existing approaches in both latency control and task usage.

Android external (shared) storage has become an essential part for current apps. When an app stores its data on external storage, a series of file operations are conducted. Compared with the file content, the information from these file operations is much limited, just telling when, where and which file operation happens. In this paper, we attempt to shed light on the great potential capability of file operations for constructing covert channels. Taking advantage of file operations on external storage, we propose File-operation-based Covert Communication Framework (FIOCC) and design three kinds of covert channels: Channel Based on Operation Type (COTY), Channel Based on Operation Target (COTA) and Channel Based on Operation Time (COTI). Meanwhile, we propose the dual-monitor mechanism to deal with the interference from other apps’ file operations. To make the file operations seem as normal as the ones of benign apps, we introduce the dynamic arrangement into the file operation generating process. we conduct both theoretical analyses and real-world experiments on these covert channels. The final results fully prove the effectiveness and applicability of the proposed file-operation-based covert channels over current Android versions.

Cooperative perception (CP) is promising to enhance environmental awareness by sharing complementary information among connected vehicles. However, the challenges of resource constraints and edge heterogeneity in vehicular networks often lead to synchronization bottlenecks and degrade CP accuracy under strict delay requirements. This paper presents RT-Cooper, a real-time CP framework that jointly orchestrates computation and communication to optimize perception accuracy with delay guarantees in heterogeneous vehicular networks. We first introduce adaptive model pruning and bandwidth allocation into the CP pipeline to enable fine-grained resource coordination. Then, we formulate the Real-time Cooperative Perception (RCP) problem, which captures coupled computation-communication delays and system-wide synchronization bottlenecks, aiming to maximize perception accuracy under hard deadlines and resource constraints. To handle the closed-box and non-convex nature of the RCP objective, we construct an interpretable feature quality model derived from pruning analysis as a surrogate for accuracy. On this basis, we develop the Two-stage Alternating Resource Optimization (TARO) algorithm, which alternates between closed-form feasibility enforcement and heterogeneity-aware feature quality refinement with provable convergence. Extensive experiments on open-source benchmark and a real-world prototype validate the effectiveness of RT-Cooper in improving perception accuracy while achieving real-time guarantees.

Expert parallelism has become the prevailing paradigm to serve Mixture-of-Experts (MoE) models. Its efficiency depends on the communication and computation latencies of the GPUs, which are linked to the placement of experts in the GPUs. Existing works for optimizing expert placement focus on leveraging past requests' expert activation patterns. However, they demonstrate deficiencies facing diverse and rapidly changing request patterns, calling for an online, proactive approach. Implementing such an approach requires addressing several challenges: the uncertainty associated with incoming requests' expert activation, the cost of expert migration, and the NP-hard complexity in optimization. Therefore, we present Director, a new distributed MoE serving system that minimizes end-to-end latency via prediction-driven, online expert placement. Director uses either a lightweight cascaded predictor or a low-bit quantized replica for expert activation patterns of incoming requests. An online migration module then enacts the changes with near-zero downtime by executing migrations in compute-bound phases, keeping disruption bounded. At its core, a relaxation-based expert placement optimizer operates under capacity constraints, runs in polynomial time, and achieves a (1+ ϵ) approximation ratio. Finally, we implement a prototype and demonstrate, through extensive experiments, a reduction in end-to-end latency of 11 ~ 55% for popular MoE models (e.g., Mistral, DeepSeek and Qwen) compared to existing work.

In Fifth Generation New Radio-based Vehicle-to-Everything (5G NR-V2X) systems, Vehicle-to-Vehicle (V2V) communication plays a vital role in improving road safety by allowing a timely exchange of situational awareness among vehicles. Accurate time synchronization is critical to ensure reliable message delivery, especially in out-of-coverage scenarios where cellular infrastructure and Global Navigation Satellite System (GNSS) services are not reliable. Although the 3rd Generation Partnership Project (3GPP) standard specifies synchronization mechanisms based on Sidelink Synchronization Block (SSB) and selection/reselection of a synchronization reference vehicle (SyncRef), these methods suffer from inherent vulnerabilities due to their dependence on broadcast transmissions and naive selection/reselection procedures. In this paper, a targeted spoofing attack that exploits this frailty is proposed under which malicious de-synchronization is introduced into the network. We conduct extensive simulations to evaluate the impact of our proposed attack. Our findings reveal a significant degradation in communication performance under the proposed attack, including increased collision probability, reduced throughput, and reduced Packet Reception Ratio (PRR). To counter this threat, we also propose a detection framework based on Channel Busy Ratio (CBR) and Channel occupancy Ratio (CR), which effectively distinguishes normal channel behaviors and synchronization-based attacks. The proposed detection method effectively enhances the robustness and security of NR-V2X systems against synchronization-related vulnerabilities. To the best of our knowledge, this is the first work that contributes to addressing the security issues of the synchronization procedure in NR-V2X sidelink communication.

Ultra-Reliable Low-Latency Communication (URLLC) is an important communication service in 5G NR, designed to support extremely delay-sensitive and ultra-reliable applications. To multiplex URLLC and Enhanced Mobile Broadband (eMBB) on the same air interface, two approaches have been proposed, namely, puncturing and reservation. While puncturing can guarantee timely URLLC delivery, it degrades eMBB goodput. On the other hand, reservation may lead to under-utilization when URLLC packet arrival becomes bursty. In this paper, we propose RaP—a novel scheduling solution that intelligently integrates both reservation and puncturing. RaP exploits the small window of URLLC’s delay tolerance by employing buffering/reservation based on a Long Short-Term Memory (LSTM)-based predictor, while performing puncturing using a deep reinforcement learning (DRL) agent when actual URLLC arrival exceeds prediction. We implement RaP with a link-level 5G NR simulator and show that RaP significantly outperforms state-of-the-art schemes and achieve a comparable performance with a ideal reservation scheme assuming perfect knowledge of URLLC packet arrival.

We study the problem of constraint-aware combinatorial multi-armed bandits (CMAB), a novel extension of combinatorial semi-bandits that maximizes total rewards while adhering to long-term cost constraints. In each round, the environment generates stochastic rewards and costs for each arm from unknown distributions. The learning agent selects a combinatorial set of arms that collectively gain rewards and incur costs, which are (partially) observed as feedback to improve future decisions. While prior work has focused on minimizing regret, it often overlooks cost constraints and diverse reward structures across applications. To address this, we introduce a general framework, constraint-aware CMAB with versatile reward functions (C2MAB-V). Unlike existing works that solve a difficult discrete constrained optimization (CO) problem to control constraint violation, we propose bypassing the original discrete CO problem and using a novel relaxation and rounding (RR) approach, which solves a relaxed continuous CO problem with improved approximation guarantees and computational efficiency. A new challenge arises since RR is originally designed for offline CO problems, and due to the flexibility of choosing the relaxation function F and the rounding procedure σ, it is unclear how to guarantee that RR can achieve low regret and low violation simultaneously. In this paper, we are the first to discover unified RR conditions and, equipped with these conditions, we prove that C2MAB-V has the following nice properties: sub-linear regret, fast diminishing violation, and good computational efficiency. We demonstrate the generality of our RR conditions by applying C2MAB-V to diverse applications with concrete (F,σ), i.e., mobile crowdsensing, network routing, and multi-LLM selection, with varying reward functions (e.g., linear, conjunctive, disjunctive, submodular) and feedback models (e.g., semi-bandit, cascading). Extensive experiments on synthetic and real-world datasets, including evaluations with nine LLMs on practical tasks, confirm that C2MAB-V outperforms baselines across these applications.

In distributed training of Large Language Models (LLMs), parallelism strategies induce iterative parameter synchronization, generating periodic, high-volume flows that persist across training iterations. Concurrently, training processes produce lightweight, bursty flows for control and coordination. The coexistence of these long-lived and bursty flows introduces substantial traffic heterogeneity, increasing pressure on the network and degrading overall training efficiency. We propose BiPlane, a behavior-aware cross-layer interconnect architecture designed to alleviate communication bottlenecks and improve training efficiency. Our key idea is to adaptively align flow characteristics with the complementary capabilities of connection-oriented and connectionless transport layers, enabling coordinated and efficient forwarding across heterogeneous layers. BiPlane proposes a two-fold design. First, BiPlane identifies periodic, connection-stable flows and maps them to connection-oriented paths to reduce communication bubbles, while routing bursty or unpredictable flows over agile connectionless fabrics. Second, BiPlane jointly models flow-level behaviors and network states in real time to inform forwarding decisions, and employs the Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO) framework for long-horizon policy learning that balances reconfiguration cost and delivery efficiency. Simulations with real LLaMA 2-70B data show that BiPlane shortens iteration time by over 2× and achieves high stability compared to traditional baselines.

Abnormal Battery Drain (ABD) is a significant yet often elusive issue for mobile apps, frequently arising from subtle network-related inefficiencies, such as redundant data transmissions or retry loops. Existing research includes static and dynamic detection methodologies, but both struggle to detect such ABDs at runtime, especially when they are triggered by network operations without direct user interaction.We present NetDrain, a lightweight runtime anomaly detection framework designed to identify and diagnose network-induced ABDs in mobile apps. NetDrain leverages thread-level CPU and network performance event counters, extracts time- and frequency-domain features, and applies an unsupervised data-driven model to detect deviations in hardware resource usage patterns. To facilitate root cause analysis, our framework correlates anomalies with specific threads and functions via circular call-graph sampling. Our evaluation shows that NetDrain successfully detects network-induced ABDs for 22 real-world apps and locates their root causes in the app code, outperforming prior research such as eDelta and eDoctor.

Routing is central to meeting the stringent, heterogeneous quality-of-service (QoS) demands of modern Internet applications. Multi-criteria routing, which computes Pareto-optimal paths over metrics such as delay and bandwidth, offers a promising solution, yet current proposals are confined to SDN designs or vector protocols. Enabling multi-criteria optimality within widely deployed link-state protocols is still an open challenge. We fill this gap with PPF (Pareto-optimal Path First), a link-state protocol supporting multi-criteria optimality with provable guarantees. PPF contributes an efficient Dijkstra-Pareto algorithm and a tag-based forwarding scheme that rides atop existing data planes (e.g., MPLS). We formally derive the algebraic conditions ensuring path- and attribute-completeness and prove routing consistency. We implement and evaluate PPF on real-world topologies using Mininet. The results show that PPF converges 95% faster while incurring up to 100× lower communication overhead than vector-based protocols. PPF also sustains lower latency and higher throughput compared to state-of-the-art QoS schemes. These results highlight PPF’s viability as a scalable, efficient, and QoS-aware multi-criteria routing solution for real-world networks.