INFOCOM 2026 Papers

Network reconfiguration has emerged as a critical task in the face of evolving and dynamic network systems, posing potential risks to the security and reliability of the network service. Current network reconfiguration frameworks focus on security requirements, using brute-force methods to search for update sequences or employing Boolean expressions to generate updates. However, these approaches consider the simulator as a black box and fall short in fine-grained control and management of the reconfiguration process that induces operator-defined priority constraints.To solve the problem, we augment security specifications with priority specifications and propose CRONUS, CounteRexample-guided cOnstraint learning for Network Update Synthesis, the first framework that efficiently synthesizes network configuration updates while meeting both security and priority requirements. Our key contribution lies in the design of learning temporal constraints from counterexamples and automatic translation of priority specifications and learned constraints into a Verilog circuit model, enabling the use of a hardware model checker to efficiently generate the update sequence.We implemented and evaluated CRONUS on real-world and synthesized topologies. The results show that CRONUS achieves a 35× speedup in synthesis time on average compared to SOTA as the specification complexity increases and solves 96% of the reconfiguration tasks in 2 minutes.

Driven by the right to be forgotten (RTBF), machine unlearning has become an essential requirement for privacy-preserving machine learning. However, its realization in decentralized federated learning (DFL) remains largely unexplored. In DFL, clients exchange local updates only with neighbors, causing model information to propagate and mix across the network. As a result, when a client requests data deletion, its influence is implicitly embedded throughout the system, making removal difficult without centralized coordination. We propose a novel certified unlearning framework for DFL based on Newton-style updates. Our approach first quantifies how a client’s data influence propagates during training. Leveraging curvature information of the loss with respect to the target data, we then construct corrective updates using Newton-style approximations. To ensure scalability, we approximate second-order information via Fisher information matrices. The resulting updates are perturbed with calibrated noise and broadcast through the network to eliminate residual influence across clients. We theoretically prove that our approach satisfies the formal definition of certified unlearning, ensuring that the unlearned model is difficult to distinguish from a retrained model without the deleted data. We also establish utility bounds showing that the unlearned model remains close to retraining from scratch. Extensive experiments across diverse decentralized settings demonstrate the effectiveness and efficiency of our framework.

Enterprises are increasingly threatened by Advanced Persistent Threats (APTs), carried out by skilled adversaries and remaining undetected for months. A promising approach to detect such intrusions is to parse logs into provenance graphs that capture data dependencies. However, a major challenge with this approach is that logs can rapidly grow to enormous sizes, imposing severe memory overhead. While existing research has introduced forensic-informed methods to reduce log size, these methods achieve modest reductions and may rely on offline processing, limiting their scalability for real-time analysis.In this work, we present Nano, a real-time log reduction approach that consolidates subject dependencies into program-specific provenance. Nano introduces two novel data structures, the profile hierarchy and the access network. Rather than preserving parent-child relationships between subjects, the profile hierarchy abstracts subject origins by capturing execution relationships between programs. For subjects created through an identical execution order of programs, the access network consolidates their activities into dependencies between executing programs and system entities, while retaining dependencies between attack activities. Our evaluation, using logs from government-agency sponsored red team exercises, demonstrates that Nano can effectively detect attacks comparable to existing rule-based intrusion detection systems, while reducing logs by up to 219 times at runtime.

With the proliferation of real-time intelligent applications, maintaining information freshness, quantified by the Age of Information (AoI), has become a critical mission for edge intelligence systems. Batch processing has become an indispensable component in the efficient scheduling of computational resources. In this work, we study the first age-minimal GPU batching problem in a multi-agent environment, driven by the coupling between users’ data generation decisions and the server’s batching policy. To tackle the scalability challenge of the curse of dimensionality, we leverage heterogeneous mean-field reinforcement learning (HMFRL). We formulate the system as a heterogeneous multi-agent Semi-Markov Decision Process (SMDP) and design a scalable, asynchronous Proximal Policy Optimization (PPO) algorithm, termed HMF-PPO. This framework approximates the complex N-agent interactions with the simplified dynamics between a representative agent and a "mean field" that characterizes the collective behavior, thereby effectively mitigating non-stationarity and ensuring scalability. Theoretical analysis reveals that the mean-field approximation error converges to zero at the rate of ${\mathcal{O}}(1/\sqrt N )$. Experimental results demonstrate that HMF-PPO can achieve up to 47% AoI reduction compared to no-batching scheduling. It also outperforms leading baselines in terms of convergence stability and scalability, and support incremental online learning for dynamic real-world edge environments.

We address the problem of orchestrating machine learning (ML) workloads, such as latency-sensitive mobile applications and intelligent radio access network (RAN) control functions, in a heterogeneous edge cloud infrastructure, subject to inference time constraints. We introduce NeuRO, a modular profiling framework that learns to approximate ML task inference time summary statistics using data collected from edge deployments. NeuRO leverages two lightweight neural embeddings: one capturing the resource footprint of an individual ML task, and one capturing the aggregate server context induced by co-located workloads. We show the potential of NeuRO by formulating a performance-aware service placement problem subject to inference time summary statistic constraints. To solve this problem efficiently at scale, we develop a heuristic algorithm that incrementally places applications across servers in a sequential manner. For each server, the constrained optimization subproblem is solved using the barrier method, ensuring tractability despite the non-linear neural network (NN) constraints. We evaluate NeuRO on heterogeneous Kubernetes clusters using realistic ML/AI workloads. Experimental results show that NeuRO not only achieves substantially higher profiling accuracy compared to existing methods, but also improves operator revenue by up to 30%, all while satisfying diverse service-level agreements (SLAs).

Verifying network configurations under all possible environments is critical for uncovering latent configuration bugs. However, due to the vastness of the environment space and the limitations of existing approaches, current tools either fail to scale or explore only a tiny fraction of the full environment space, missing many potential issues. In this paper, we show that it is unnecessary to explore the full environment space. Instead, we introduce an effective reduction of the environment space, under which the network preserves all routing and forwarding behaviors exhibited under the full environment space. We develop a theory of network behavior equivalence between the full and effectively reduced environment spaces, and propose an efficient algorithm to compute such an effective reduction. We implement our approach in a tool called Bulwark. Experiments on both real-world and synthetic networks show that Bulwark uncovers numerous new misconfigurations and delivers orders-of-magnitude speedups over state-of-the-art verifiers.

Large language models (LLMs) have revolutionized natural language processing, yet their high computational demands pose significant challenges for real-time inference, especially in multi-user server speculative decoding and resource-constrained environments. Speculative decoding has emerged as a promising technique to accelerate LLM inference by using lightweight draft models to generate candidate tokens, which are subsequently verified by a larger, more accurate model. However, ensuring both high goodput (the effective rate of accepted tokens) and fairness across multiple draft servers cooperating with a central verification server remains an open challenge. This paper introduces GOODSPEED, a novel distributed inference framework that optimizes goodput through adaptive speculative decoding. GOODSPEED employs a central verification server that coordinates a set of heterogeneous draft servers, each running a small language model to generate speculative tokens. To manage resource allocation effectively, GOODSPEED incorporates a gradient scheduling algorithm that dynamically assigns token verification tasks, maximizing a logarithmic utility function to ensure proportional fairness across servers. By processing speculative outputs from all draft servers in parallel, the framework enables efficient collaboration between the verification server and distributed draft generators, streamlining both latency and throughput. Through rigorous fluid sample path analysis, we show that GOODSPEED converges to the optimal goodput allocation in steady-state conditions and maintains near-optimal performance with provably bounded error under dynamic workloads. These results demonstrate that GOODSPEED provides a scalable, fair and efficient solution for multi-server speculative decoding in distributed LLM inference systems.

Overlapping communication and computation operators is a common practice to hide communication overheads, accelerating large language models (LLMs) training on GPU clusters. Existing systems achieve this through either intra-operator fusion (IntraFusion), which packs operators into a single large kernel, or inter-operator decomposition (InterDecom), which splits a tensor into multiple parts for pipelined execution. However, current IntraFusion methods underutilize network topology, causing suboptimal bandwidth usage on multi-GPU systems, while InterDecom struggles to determine the optimal number of decomposed parts for peak performance. To address these issues, we introduce Compass, which employs systematic optimization and comprehensive modeling. First, we design a novel IntraFusion algorithm leveraging double-ring communications to maximize bandwidth utilization in hybrid NVLink-PCIe systems, achieving 1.5x-2.5x speedups. Second, we develop a decomposition model that mathematically derives the optimal tensor decomposition degree for InterDecom, improving performance by up to 1.3x. Finally, we develop a unified performance framework that accurately determines the best strategy for different scenarios. We validate Compass through extensive evaluation across 288 configurations and end-to-end experiments on real-world applications. The results demonstrate that Compass consistently selects the optimal strategy, achieving up to a 1.42x end-to-end speedup compared to the Megatron-LM baseline.

Federated Learning (FL) enables collaborative model training without exposing private data, but remains vulnerable to backdoor attacks, where malicious clients inject backdoor updates into the global model. Detecting such attacks is challenging due to the noisy and heterogeneous nature of benign updates obscuring backdoor patterns. To this end, we propose Peeler, a lightweight backdoor defense framework that accurately identifies backdoor by dynamically isolating normal model weights orthogonal to malicious ones within client-submitted updates. Peeler employs circuit discovery to prune subnetworks associated with the main task, and utilizes gradient-based layer selection to eliminate layers that are non-critical for backdoor objective adaptation. The remaining parameters are then flattened into feature vectors, enabling distance-based anomaly detection across client-submitted model updates. To validate Peeler, we perform a Neural Tangent Kernel (NTK)-based analysis, showing that Peeler effectively retains backdoor-relevant weights while filtering out benign ones. Experiments across diverse scenarios and recent defense benchmarks demonstrate the superiority of Peeler. Peeler reduces the Attack Success Rate (ASR) to 2.79% on average, significantly outperforming the prior state-of-the-art defense (11.48% for FLAME). Even under highly heterogeneous data distributions (with Dirichlet parameter α = 0.3), Peeler achieves an ASR of 10.4%, compared to the prior state-of-the-art defense (27.3% for FLTracer).

The current programming and execution models of programmable switches are overly specialized, requiring ad hoc design choices to overcome rigid hardware constraints when implementing particular computations. We attribute this limitation to the absence of Turing completeness, which decouples computation from its execution mechanism. To address this, we propose Turing-complete virtual machines on programmable switches. Our design enables programs written in high-level general-purpose languages, such as C, to run on the switch data plane with execution dynamically determined at runtime. To this end, we first demonstrate the feasibility of universal computation by implementing a Turing machine and then extend it to a register machine with a C compiler targeting it. We further enhance this design into the Anomalocaris machine, optimized for performance through parallel instruction execution and efficient memory access, tailored to a state-of-the-art programmable switch architecture. Our evaluation demonstrates the performance and applicability of Anomalocaris across synthetic workloads and real-world network functions.

Next-generation wireless networks, including satellite-to-Open RAN systems, demand agile and intelligent resource management capable of handling dynamic multi-user interference under stochastic quality of service constraints. This paper introduces DIFFRACT, a neuralized utility maximization framework that leverages differentiable programming to integrate deep learning with optimization in wireless networks. Central to our approach is the exploitation of the mathematical structure of standard interference functions, which are foundational in wireless power control. By developing a duality theory for these functions, we map iterative interference management algorithms into differentiable neural network architectures via algorithm unrolling. This enables distributed, end-to-end gradient-based learning at the network edge, supporting real-time adaptation to interference in both terrestrial and non-terrestrial environments. DIFFRACT allows for scalable and robust utility maximization by modeling complex channel dynamics and leveraging the expressiveness of differentiable models. Experimental results confirm the framework’s theoretical soundness and practical effectiveness for next-generation wireless systems.

Fine-tuning is essential to adapt general-purpose large language models (LLMs) to domain-specific tasks. As a privacy-preserving framework to leverage decentralized data for collaborative model training, Federated Learning (FL) is gaining popularity in LLM fine-tuning, but remains challenging due to the high cost of transmitting full model parameters and computing full gradients on resource-constrained clients. While Parameter-Efficient Fine-Tuning (PEFT) methods are widely used in FL to reduce communication and memory costs, they often sacrifice model performance compared to FFT. This paper proposes FedKRSO (Federated K-Seed Random Subspace Optimization), a novel method that enables communication and memory efficient FFT of LLMs in federated settings. In FedKRSO, clients update the model within a shared set of random low-dimension subspaces generated by the server to save memory usage. Furthermore, instead of transmitting full model parameters in each FL round, clients send only the model update accumulators along the subspaces to the server, enabling efficient global model aggregation and dissemination. By using these strategies, FedKRSO can substantially reduce communication and memory overhead while overcoming the performance limitations of PEFT, closely approximating the performance of federated FFT. The convergence properties of FedKRSO are analyzed rigorously under general FL settings. Extensive experiments on the GLUE benchmark across diverse FL scenarios demonstrate that FedKRSO achieves both superior performance and low communication and memory overhead, paving the way towards on federated LLM fine-tuning at the resource-constrained edge.

Real-time video conferencing is increasingly prevalent in everyday life, yet it faces growing threats from forgery attacks where synthetic or replayed videos are injected into live calls, potentially leading to serious privacy breaches or financial losses. Unfortunately, existing approaches remain inadequate for detecting such attacks in a reliable, efficient, and user-friendly manner. In light of this, we propose EchoFence, a non-intrusive, lightweight, and robust framework for authenticating live video streams. EchoFence actively emits imperceptible ultrasonic signals during video conferencing, which physically interact with user’s natural facial and body movements and are captured by the microphone. Motion-related features are then extracted from both the ultrasonic responses and the video frames, and a training-free cross-modal verification strategy is employed to assess their temporal coherence. Significant misalignment between the two modalities is taken as strong evidence of forgery. Additionally, each ultrasonic signal carries a random credential via frequency modulation, which is validated through template-based matching, preventing tampering attempts involving ultrasound replay or removal. Extensive experiments show that EchoFence effectively detects three representative types of video forgeries with over 94% accuracy, and remains robust under diverse conditions, making it a practical solution for trustworthy video conferencing.

Low Earth Orbit (LEO) mega-constellations promise seamless global service, yet satellites racing overhead force handovers every few minutes, disrupting service for billions of users worldwide. Current systems maintain exclusive single-path connections despite modern phased arrays’ ability to track multiple satellites simultaneously, accepting these disruptions as unavoidable and merely minimizing instead of eliminating them. We propose PMPS (Predictive Multi-Path Scheduling), designed to achieve handover-free satellite communications through proactive continuous multi-path service. PMPS eliminates the reactive single-path service interruption through three integrated designs: (i) introducing multiplicative composite weights combining visibility time, signal quality, and capacity factors orchestrates continuous traffic redistribution, ensuring gradual and seamless packet migration across all active satellites. (ii) proving optimal multi-satellite packet scheduling is NP-complete, then developing polynomial-time approximation algorithms achieves provably near-optimal performance within real-time constraints. (iii) eliminating obstruction-induced packet loss through anticipatory selective redundancy protects vulnerable connections. As satellites approach the horizon becoming vulnerable to obstruction, their declining weights have shifted most traffic to higher satellites. Redundantly transmitting only the minimal remaining packets through stable satellites ensures zero loss if obstruction occurs. Evaluation on Starlink and OneWeb constellations demonstrates PMPS enables handover-free LEO communications with minimal redundancy overhead.

Sharding is a key approach to improving blockchain scalability for parallel transaction processing. Among sharding strategies, state sharding splits the ledger across multiple shards, allowing each validator to process only a subset of transactions; however, maintaining consistency for cross-shard transactions remains challenging. Most existing sharding systems lack scalability due to reliance on Merkle proofs, centralized intermediaries, or redundant ledger storage. In this paper, we propose ShardTree, a cross-shard protocol that enables high throughput and low confirmation latency for sharded blockchains without these limitations. ShardTree leverages a multi-party virtual payment channel (MPC) to efficiently manage cross-shard transaction processing in batches. We first design an algorithm to select validators to form an MPC with the maximum capacity over a payment channel (PC) path. Based on this, ShardTree constructs an MPC over a PC tree. We also design a rollback scheme to ensure the atomicity of cross-shard transactions. We theoretically prove that ShardTree guarantees security in the presence of Byzantine adversaries. Finally, we implement ShardTree and assess its performance using real-world Ethereum transactions. Evaluation results demonstrate that ShardTree efficiently processes cross-shard transactions and outperforms state-of-the-art protocols in terms of transaction throughput and confirmation latency.

Retrieval Augmented Generation (RAG) greatly mitigates hallucinations in large language models by grounding responses in external documents, yet its adoption on mobile devices is hindered by two intertwined concerns: (i) security, because poisoned passages cached by adversaries can drive the model to generate harmful outputs, and (ii) efficiency, because deeper retrieval and wider uplinks consume limited battery and spectrum. Existing RAG security studies overlook bandwidth and power budgets, whereas mobile edge offloading literature assumes benign knowledge bases. Bridging this gap, in this paper we introduce the Security Cost Ratio (SCR) as a unified metric that rewards lower attack-success rate per joule expended, and we design an optimization problem to jointly control retrieval depth, bandwidth and transmit power in the mobile edge computing environment. To tackle the non-convex optimization problem, we derive closed-form KKT conditions and an epigraph transformation that attain the global optimum for retrieval depth. We then utilize a quadratic-transform fractional programming technique in an alternating framework that converges for bandwidth and transmit power allocation. Simulations across various system parameters and environmental settings demonstrate that the proposed algorithm achieves up to 1.8× higher SCR and 50% lower attack success rate, while maintaining rapid convergence.

Industrial quality inspection faces a fundamental trade-off between large-field-of-view (FOV) coverage and high precision. As a representative case, fabric density, a critical textile quality metric, directly affects fabric’s performance, with uncorrected deviations causing large-scale defects. Existing methods rely on close-range imaging to capture fine fabric details, failing to achieve both a large FOV and high precision simultaneously. In this paper, we propose MoiréDensimeter, a high-precision density measurement solution for real-time production, utilizing low-cost COTS cameras to enable large-FOV monitoring. By innovatively introducing moiré patterns as a visual medium, we transform the micron-scale high-frequency fabric textures captured at long distances into recognizable low-frequency moiré pattern, enabling high-precision density measurement in a large FOV. To bridge moiré features with fabric density, we develop a mathematical model and propose a clearness-adaptive density scaling strategy to extend the measurement range. To address complex industrial environmental interference, we propose a Demosaicing Feature Map (DFM) method based on high-frequency characteristics in moiré pattern generation, enabling precise fabric detection in large field of view images. Extensive experimental results show that MoiréDensimeter achieves a large-FOV measurement of sub-millimeter density at a maximum distance of 6.5 meters, with an average micron-level relative error of 0.245%.

Indoor Positioning System (IPS) is essential for enabling location-based services in complex environments such as malls and airports. However, traditional methods struggle with cost and scalability, requiring either expensive infrastructure or labor-intensive data collection. We propose a Wi-Fi Aware-based IPS, called AIPS, a fully infrastructure-free localization system that combines Wi-Fi Aware-based peer-to-peer ranging with an inertial measurement unit, global navigation satellite system, and AI-driven indoor/outdoor detection. The system opportunistically collaborates with nearby mobile devices and leverages transaction events from mobile payment to build self-generating radio maps. To ensure the robustness of localization, AIPS adaptively switches between advanced least square and extended particle filters depending on the number of nearby mobile devices. AIPS achieves 2.62 m accuracy with a minimal battery energy cost of 0.2% per hour, offering a self-evolving and commercially viable solution for next-generation positioning.