INFOCOM 2026 Papers

Leader election serves a well-defined role in leader-based Byzantine Fault Tolerant (BFT) protocols. Existing reputation-based leader election frameworks for partially synchronous BFTs suffer from either protocol-specific proofs, narrow applicability, or unbounded recovery after network stabilization, leaving an open problem. This paper presents a novel protocol-independent abstraction formalizing generic correctness properties and effectiveness guarantees for leader election under partial synchrony, enabling protocol-independent analysis and design. Building on this, we design the Sliding Window Leader Election (SWLE) mechanism. SWLE dynamically adjusts leader nominations via consensus-behavior-based reputation scores, enforcing Byzantine-cost amplification. We demonstrate SWLE introduces minimal extra overhead to the base protocol and prove it satisfies all abstraction properties and provides superior effectiveness. We show, with a 16-server deployment across 4 different regions in northern China, SWLE achieves up to 4.2x higher throughput, 75% lower latency and 27% Byzantine leader frequency compared to the state-of-the-art solution under common Byzantine faults, while maintaining efficiency in fault-free scenarios.

Edge computing enables resource-constrained devices to execute machine learning applications via task offloading. To this aim, radio access network (RAN) slicing is instrumental to provide the necessary network resources. However, current RAN slicing approaches rely on static computing models, thereby constraining their ability to leverage the dynamic semantic data representation capabilities enabled by recent neural architectures. In this paper, we propose OMNIS, a semantic RAN slicing framework for edge computing built on dynamic split neural models. OMNIS embeds a dynamic form of neural compression paired with adaptive data encoding for task offloading, enabling flexible communication payload and computing options. We explicitly study the interplay between neural compression and information quantization in computer vision tasks and design a novel "Box" quantization scheme that improves resiliency to bit errors as a function of compression rate. Considering partial observability and differing objectives of mobile devices and the edge server, we formulate neural gate control and resource slicing optimization problems and solve them via multi-agent contextual multi-armed bandits and convex optimization algorithms. Experimental results show that OMNIS improves inference accuracy by up to 22.85% and reduces quality of service violations by up to 10×. The evaluation code is available at https://github.com/qlt315/OMNIS.

Network Digital Twins (NDTs) are reshaping the network research landscape by enabling real-time predictive modeling and analysis of complex networked systems. Despite their potential, practical deployment of NDTs is hindered by growing divergence between the physical network and its digital replica, driven by insufficient or delayed synchronization, and the high communication overhead from continuous calibration, often worsened by inefficient data representation. To address these issues, we propose Twintinuum, a novel uncertainty-guided self-calibration mechanism that enables scalable twinning with low communication overhead. Our approach dynamically adapts the synchronization based on estimated model uncertainty, enabling both short-term responsiveness and long-term consistency. Moreover, the proposed framework supports plug-and-play calibration, allowing the digital twins to autonomously align with physical dynamics without extensive manual tuning or retraining. Extensive experiments on wireless signal datasets and diverse base models demonstrate that our method achieves superior signal quality estimation with significantly less data exchange.

Emerging high-quality real-time applications require consistently low latency, which is often disrupted by latency spikes. We identify the reason as the mismatch between the abrupt bandwidth reallocation on routers and gradual sending rate reaction of congestion control. For example, when a burst of new flows arrives, queue schedulers such as fair queueing immediately reallocate the bandwidth for existing and new flows. However, the flow's sending rate, determined by the congestion control algorithm (CCA), needs several RTTs to converge to the new available bandwidth, during which severe stalls occur. This has been increasingly critical with the demand on consistent low latency. In this paper, we present Confucius, a practical queue management scheme that reallocate the bandwidth for flows following CCA's reaction. Confucius slows down bandwidth adjustment to match the reaction of congestion control, so that the end host can reduce the sending rate without overshooting the network. Confucius is designed for offering real-time flows with consistently low latency regardless of uncertain competition. Experiments show that Confucius reduces the stall duration by more than 50% against existing practical schemes, while competing flows also fairly enjoy on-par performance.Available at: https://github.com/hkust-spark/confucius-qdisc

In a federated learning (FL) system, many devices, such as smartphones, are often undependable (e.g., frequently disconnected from WiFi) during training. Existing FL frameworks always assume a dependable environment and simply exclude undependable devices from training, leading to poor model performance and resource wastage. In this paper, we propose FLUDE to effectively deal with undependable environments. First, FLUDE assesses the dependability of devices based on the probability distribution of their historical behaviors (e.g., the likelihood of successfully completing training). Based on this assessment, FLUDE adaptively selects devices with high dependability for training. To mitigate resource wastage during the training phase, FLUDE maintains a model cache on each device, aiming to preserve the latest training state for later use in case local training on an undependable device is interrupted. Moreover, FLUDE proposes a staleness-aware strategy to judiciously distribute the global model to a subset of devices, thus significantly reducing resource wastage while maintaining model performance. We have implemented FLUDE on two physical platforms with 120 smartphones and NVIDIA Jetson devices. Extensive experiments show that FLUDE effectively improves model performance and resource efficiency of FL in undependable environments.

Low Earth Orbit satellites play a vital role in Earth observation and remote sensing missions, with their satellite-to-ground downlinks responsible for transmitting substantial volumes of imagery and data. However, the inherent dynamism and instability of satellite-to-ground links present challenges to efficient data transmission. While existing research often concentrates on predicting physical layer parameters, our empirical measurements reveal that solely relying on physical layer predictions is insufficient to reflect application layer throughput accurately. To address this, we propose Satformer, a spatio-temporal prediction model, and based on it, the Spatio-Temporal cross-layer Adaptive Rate control (STAR) mechanism to achieve efficient satellite-to-ground downlink transmission. To bridge the gap between physical and application layers, Satformer employs tailored Sat-Embedding and a spatio-temporal attention mechanism that explicitly correlates satellite relative position data and weather data with physical layer data to predict application layer throughput. Experimental results show that Satformer improves application layer throughput prediction accuracy by up to 33.45% compared to baseline models. Building on this, the STAR mechanism increases satellite-to-ground data transmission throughput by an average of 12.12%, while reducing median end-to-end latency by up to 43.3%.

Trajectory datasets collected by network operators and service providers offer detailed information about individual mobility and have wide application in business and research. However, managing such data raises privacy risks, as the unique movement patterns of individuals pose significant re-identification risks and make common countermeasures like pseudonymization ineffective. The privacy-preserving data publishing (PPDP) of trajectory datasets that maintains post-anonymization accuracy and truthfulness is an open problem -especially for large datasets with millions of records like those gathered by major actors in the telco ecosystem. We close this gap with k-scale, a framework that implements k-anonymity in massive mobile user trajectory datasets, removing uniqueness while safeguarding accuracy at the record level. Not only k-scale is the first model capable of scaling k-anonymization to a dataset of one million trajectories, but it does so while also outperforming state-of-the-art methods for trajectory data publishing in terms of preserved data quality, which we prove in real-world massive datasets and applications.

Network operators utilize in-network bandwidth allocation strategies to ensure fairness among competing flows. Existing approaches mainly rely on two mechanisms deployed in switches: packet scheduling algorithms and probabilistic packet dropping. However, their effectiveness is constrained by the limited number of physical priority queues and the hardware overhead of arithmetic operations. The former leads to high packet loss, resulting in low link utilization, while the latter impairs bandwidth allocation accuracy.In this paper, we propose AGILE, a novel bandwidth allocation strategy that enhances network fairness while maintaining high link utilization. Despite the limited flow visibility in switches, AGILE performs real-time global flow state tracking and optimal per-flow fair share computation to achieve max-min fairness, dynamically allocating residual bandwidth to maximize link utilization. We implement a prototype of AGILE on Intel Tofino switches. Our experimental results show that AGILE attains an average Jain’s Fairness Index (JFI) of 0.996. Compared to state-of-the-art approaches, including Cebinae and AHAB, AGILE reduces relative bandwidth allocation errors by 45.7%–86.3%, improves link utilization by 5%, and decreases link utilization fluctuations by 37%–75%.

This paper presents a practical design and simulation framework for a reconfigurable intelligent surface (RIS) that enables high-precision signal cancellation in a three-dimensional space. The system targets airborne ADS-B transmissions at 1090 MHz—now a major source of interference for radio astronomy observatories. Unlike conventional RIS architectures that support only phase or amplitude control, we introduce a novel unit cell that integrates independently biased PIN and varactor diodes to achieve full two-dimensional tunability of the reflection coefficient. The unit cell is modeled using nonlinear circuit simulations in OrCAD PSpice and validated via full-wave electromagnetic analysis in Ansys HFSS. A tailored admittance-matching method translates target reflection coefficients into feasible RIS configurations. Simulations of a 64×64 RIS array show that more than 97% of the 2993 evaluated directions of arrival achieve sub-0 dB residual power at the receiver, demonstrating robust suppression of mobile interference. This work establishes the feasibility of RIS-based spatial RFI cancellation at a fixed frequency, with implications for spectrum coexistence, secure communications, and ultra-sensitive sensing platforms.

Large Language Models (LLMs) are revolutionizing how users interact with information systems, yet their high inference cost poses serious scalability and sustainability challenges. Caching inference responses, allowing them to be retrieved without another forward pass through the LLM, has emerged as one possible solution. Traditional exact-match caching, however, overlooks the semantic similarity between queries, leading to unnecessary recomputation. Semantic caching addresses this by retrieving responses based on semantic similarity, but introduces a fundamentally different cache eviction problem: one must account for mismatch costs between incoming queries and cached responses. Moreover, key system parameters, such as query arrival probabilities and serving costs, are often unknown and must be learned over time. Existing semantic caching methods are largely ad-hoc, lacking theoretical foundations and unable to adapt to real-world uncertainty. In this paper, we present a principled, learning-based framework for semantic cache eviction under unknown query and cost distributions. We formulate both offline optimization and online learning variants of the problem, and develop provably efficient algorithms with state-of-the-art guarantees. We also evaluate our framework on a synthetic dataset, showing that our proposed algorithms perform matching or superior performance compared with baselines.

Digital Twin (DT) creates a twin world of physical systems, enabling intelligent decision-making through predictive simulation and data analysis. Its effectiveness relies on precise synchronization between physical and twin worlds through up-to-date data transmission, which is challenging in large-scale deployments. This work explores a large-scale DT system where heterogeneous physical entities share a common network resource to update state. We quantify synchronization error using Version Age of Information (VAoI), tracking the number of unsynchronized updates. Baseline policies suffer from either poor synchronization or high communication overhead. As such, we propose a low-overhead Threshold-based VAoI-aware Power-of-d (TV-PoD) policy, wherein the scheduler randomly samples d entities, among which selects those with VAoI exceeding the threshold, and schedules the one with the maximum VAoI for state update. Analyzing TV-PoD is challenging due to the coupled VAoI processes across entities. In the large-system limit, such coupling becomes negligible, allowing us to develop a VAoI-oriented mean-field framework that characterizes system-level VAoI dynamics in heterogeneous settings, and quantifies TV-PoD’s performance. Simulations validate our theoretical results and demonstrate that TV-PoD enables substantial overhead reduction while preserving synchronization quality. DT prototyping confirms the practical feasibility of our policy. To facilitate research in DT networking, we open-source the prototype [1].

Federated Unlearning (FUL) focuses on client data and computing power to offer a privacy-preserving solution. However, high computational demands, complex incentive mechanisms, and disparities in client-side computing power often lead to long times and higher costs. To address these challenges, many existing methods rely on server-side knowledge distillation that solely removes the updates of the target client, overlooking the privacy embedded in the contributions of other clients, which can lead to privacy leakage. In this work, we introduce DPUL, a novel server-side unlearning method that deeply unlearns all influential weights to prevent privacy pitfalls. Our approach comprises three components: (i) identifying high-weight parameters by filtering client update magnitudes, and rolling them back to ensure deep removal. (ii) leveraging the variational autoencoder (VAE) to reconstruct and eliminate low-weight parameters. (iii) utilizing a projection-based technique to recover the model. Experimental results on four datasets demonstrate that DPUL surpasses state-of-the-art baselines, providing a 1%-5% improvement in accuracy and up to 12x reduction in time cost.

Modern data center networks underpin latency-critical services such as cloud computing, AI workloads and emerging 6G applications, yet even brief switch or link failures can cascade into severe service disruptions. Existing fault detection systems face limitations: control plane centric methods incur high latency (seconds to minutes) and suffer control channel overload, while data-plane-centric approaches often rely on intrusive probing, limiting scalability. We present MPulse, a multi-dimensional fault detection and recovery framework. Our primary idea is to differentiate the roles of monitoring switches and transmission switches, shifting fault management to the data plane through two key innovations: collaborative multidimensional monitoring and a decentralized voting mechanism. Following this, we enable switches to collaboratively collect real-time status, historical data, and topological state through periodic information exchange, allowing each node to aggregate global multi-dimensional insights. Then, a distributed voting process is designed to allow switches to independently determine fault types (link/switch failure) and locations based on integrated data dimensions, eliminating dependencies on centralized control. Implemented on Intel Tofino switches with minimal overhead, MPulse achieves sub-millisecond fault localization (<1 ms), 324.8× faster detection than existing solutions, and near-100% accuracy under congestion. Experimental results validate its effectiveness and low operational costs.

Automatic speech enhancement and separation are critical for improving speech quality and intelligibility in noisy, multi-speaker environments, with applications in hearing aids, automatic speech recognition (ASR), and voice-controlled systems. However, most existing methods rely on task-specific models designed for a fixed number of speakers—such as one model for single-speaker enhancement and others for two- or three-speaker separation—requiring explicit prior knowledge of speaker count. This design paradigm limits flexibility and scalability in real-world scenarios where the number of speakers is often unknown and dynamic. To overcome this limitation, we introduce Ra-dioSEP, a unified mmWave-audio multimodal framework that performs both speech enhancement and separation without prior knowledge of speaker count. By leveraging mmWave radar sensing, RadioSEP detects, localizes, and profiles multiple speakers, enabling dynamic estimation of speaker count and extraction of speaker-specific physical features. These features are fused with noisy audio and processed by a single deep neural network that adapts to the current speaker configuration, generating clean speech streams accordingly. Extensive experiments show that RadioSEP consistently outperforms state-of-the-art methods in both speech enhancement and multi-speaker separation tasks, while offering significantly improved generalization and adaptability.

Cellular networks are comprised of software-based entities, with main functions encapsulated as Virtual Network Functions (VNFs) deployed on Commercial-off-the-Shelf (COTS) hardware. As a key enabler of 5G, network slicing offers logically isolated Quality of Service (QoS) for diverse use cases. With the transition to cloud-native infrastructures, optimizing network slice placement across multi-cloud environments remains challenging due to heterogeneous resource capabilities and varying slice-specific demands. This paper presents SlicePilot, a modular framework that enables autonomous and near-optimal VNF placement using a disaggregated Multi-Agent Reinforcement Learning (MARL) approach. SlicePilot collects real-world traffic profiles to estimate resource needs for each slice type. These estimates guide a MARL-based scheduler that minimizes deployment costs while satisfying QoS constraints. We evaluate SlicePilot on a multi-cloud testbed and demonstrate a 19x speed-up over combinatorial optimization methods, while keeping deployment costs within 7.8% of the optimal. Although SlicePilot results in 2.42x more QoS violations under high-load conditions, this trade-off is offset by faster decision-making and reduced computational overhead. Overall, SlicePilot delivers a scalable, cost-efficient solution for network slice placement, making it suitable for real-time deployments where responsiveness and efficiency are critical.

Federated Learning faces significant challenges arising from two critical uncertainties: the validity of a client’s participation, which can be compromised by network and system heterogeneity, and the utility of the data contributed by each client, which varies due to heterogeneous statistical data. Traditional client selection methods often treat these uncertainties as a whole, leading to suboptimal performance. To address this issue, we propose FedSUV, an innovative client selection framework that decouples validity and utility uncertainties. FedSUV approaches client selection from a multi-objective optimization perspective, employing advanced bandit algorithms: a confidence bound-based linear contextual bandit for assessing validity and a Gaussian Process bandit for evaluating utility. We validate the effectiveness of FedSUV through both theoretical analysis and large-scale experiments conducted within our physical cluster.

We study a sampling problem, where the sampler chooses a time to take a sample of the Ornstein Uhlenbeck (OU) process and transmits the sample to a remote estimator over an unreliable channel. The goal is to design these sampling times to minimize the long-term average mean square estimation error (MSE). Our problem strictly generalizes the previous study from the Wiener process to the OU process and poses significant challenges. First, the OU process includes an additional drift term, complicating the reward function. Then, under an unreliable channel, computing the action value functions needs iterations. These make it hard to compute an optimal value function via value iteration. Moreover, the drift term makes the third derivative of the reward function nonlinear, making it complicated to validate the optimality. Despite the challenges, through thorough calculations and a convexity assumption, we are still able to obtain an optimal sampling policy that remains a simple threshold structure with low complexity. In addition, we add a sampling rate constraint that is not considered in the previous study, and provide an exact optimal sampling policy. A counterintuitive and different result is that the zero-wait policy is optimal when the channel is highly unreliable.

Modern HTTPS mechanisms such as Encrypted Client Hello (ECH) and encrypted DNS improve privacy but remain vulnerable to website fingerprinting (WF) attacks, where adversaries infer visited sites from encrypted traffic patterns. Existing WF methods rely on supervised learning with site-specific labeled traces, which limits scalability and fails to handle previously unseen websites. We address these limitations by reformulating WF as a zero-shot cross-modal retrieval problem and introducing STAR. STAR learns a joint embedding space for encrypted traffic traces and crawl-time logic profiles using a dual-encoder architecture. Trained on 150K automatically collected traffic–logic pairs with contrastive and consistency objectives and structure-aware augmentation, STAR retrieves the most semantically aligned profile for a trace without requiring target-side traffic during training. Experiments on 1,600 unseen websites show that STAR achieves 87.9% top-1 accuracy and 0.963 AUC in open-world detection, outperforming supervised and few-shot baselines. Adding an Adapter with only four labeled traces per site further boosts top-5 accuracy to 98.8%. Our analysis reveals intrinsic semantic–traffic alignment in modern web protocols, identifying semantic leakage as the dominant privacy risk in encrypted HTTPS traffic. We release STAR’s datasets and code to support reproducibility and future research1.