INFOCOM 2026 Papers

Traffic classification on programmable data plane holds great promise for line-rate processing, with methods evolving from per-packet to flow-level analysis for higher accuracy. However, a trade-off between accuracy and efficiency persists. Statistical feature-based methods align with hardware constraints but often exhibit limited accuracy, while online deep learning methods using packet sequential features achieve superior accuracy but require substantial computational resources. This paper presents Synecdoche, the first traffic classification framework that successfully deploys packet sequential features on a programmable data plane via pattern matching, achieving both high accuracy and efficiency. Our key insight is that discriminative information concentrates in short sub-sequences—termed Key Segments—that serve as compact traffic features for efficient data plane matching. Synecdoche employs an "offline discovery, online matching" paradigm: deep learning models automatically discover Key Segment patterns offline, which are then compiled into optimized table entries for direct data plane matching. Extensive experiments demonstrate Synecdoche’s superior accuracy, improving F1-scores by up to 26.4% against statistical methods and 18.3% against online deep learning methods, while reducing latency by 13.0% and achieving 79.2% reduction in SRAM usage.

Oblivious load-balancing in networks involves routing traffic from sources to destinations using predetermined routes independent of the traffic, so that the maximum load on any link in the network is minimized. We investigate oblivious load-balancing schemes for a N × N torus network under sparse traffic where there are at most k active source-destination pairs. We are motivated by the problem of load-balancing in large-scale LEO satellite networks, which can be modeled as a torus, where the traffic is known to be sparse and localized to certain hotspot areas. We formulate the problem as a linear program and show that no oblivious routing scheme can achieve a worst-case load lower than approximately $\frac{{\sqrt {2k} }}{4}$ when 12/2 and $\frac{N}{4}$ when N2/2 ≤ k ≤ N2. Moreover, we demonstrate that the celebrated Valiant Load Balancing scheme is suboptimal under sparse traffic and construct an optimal oblivious load-balancing scheme that achieves the lower bound. Further, we discover a $\sqrt 2 $ multiplicative gap between the worst-case load of a non-oblivious routing and the worst-case load of any oblivious routing. The results can also be extended to general N × M tori with unequal link capacities along the vertical and horizontal directions.

With mobile video calls now ubiquitous, ensuring seamless video-based real-time communications (RTC) remains a critical challenge for 5G operators. Despite abundant 5G bandwidth, video calls frequently experience low quality and unacceptable latency during channel fluctuations—not due to bandwidth limitations or congestion, but because of non-congestive delays in the radio access network (RAN). These delays stem from general-purpose RAN transmission procedures that prioritize radio resource efficiency over application latency through reactive scheduling and static timer-based retransmissions. Existing solutions largely address congestion-induced delays or sacrifice spectral efficiency to mitigate non-congestive delays.To address this, we present PAVE, a novel RAN-side solution that breaks the fundamental tradeoff between spectral efficiency and RTC latency. Our key insight is that RTC traffic exhibits distinct characteristics—periodic traffic generation and deadline-driven urgency—that can be leveraged to optimize RTC quality of experience (QoE) without sacrificing efficiency. For practical deployment, PAVE extracts these characteristics at the RAN, enables resource preallocation that goes beyond strictly periodic traffic patterns, and incorporates a selective retransmission and skip mechanism that maintains high spectral efficiency while strategically leveraging application-layer recovery. Implemented in an Open-RAN compliant RAN Intelligent Controller, PAVE improves tail frame rates by 1.8× and reduces video stalls by 94% in real-world evaluations.

Synthetic traffic generation is a fundamental technique for evaluating system performance and security resilience. However, existing approaches fail to capture the complex, interactive traffic patterns of modern applications. While recent deep learning models can synthesize individual traffic flows with high fidelity, they are fundamentally restricted to these isolated behaviors. They cannot reproduce the system-level interactions among multiple nodes, due to the state space of multi-node systems, which grows exponentially with the number of participants and renders direct modeling computationally intractable. To break this scalability barrier, we introduce NEST, a node-interactive generative emulation framework that enables multi-node synthetic traffic generation. Its key innovation circumvents exponential complexity by decomposing the problem: rather than modeling the global network, a generative model learns each node’s local behavior, which a lightweight scheduler then orchestrates into coherent, interactive traffic. Evaluations show that NEST not only achieves high statistical fidelity but, for the first time, successfully reconstructs the multi-node interaction graphs of real-world applications. These generated graphs replicate the complex dependency structures of real traffic with an average similarity of 97.7%, a task fundamentally unattainable by prior single-flow models.

As the push toward fully autonomous 5G and future 6G networks accelerates, Deep Reinforcement Learning (DRL) has emerged as a cornerstone of intelligent decision-making, enabling real-time adaptability and self-optimization. However, this promise is increasingly overshadowed by a critical and underexamined risk: DRL Drift, which refers to sudden and often opaque degradation in agent performance after deployment. This phenomenon jeopardizes the reliability and trustworthiness of DRL systems operating in dynamic, real-world telecom environments. Despite growing adoption of DRL in both research and industry, the issue of DRL Drift remains largely overlooked in major telecommunications standards such as the 3rd Generation Partnership Project (3GPP) and the European Telecommunications Standards Institute (ETSI). To address this challenge, we propose a novel Zero-Touch Drift Management framework, developed in alignment with the ETSI closed-loop reference architecture. At its core lies the Composite Drift Index, a unified, domain-agnostic metric that combines key performance indicators, state-action transitions, and uncertainty estimation to enable proactive detection of degradation. Extensive evaluations in a representative network-slicing scenario demonstrate up to 23.4% higher detection accuracy than baseline methods, with strong generalization across diverse DRL approaches. This work offers the first standards-aligned solution to enhance DRL resilience against drift in next-generation networks.

Recent breakthroughs in large language models have enabled various generative tasks on a single model. Real-world services powered by the LLM often concurrently support latency-critical requests for interactive applications (e.g., question-answering systems, referred to as real-time or RT requests) and throughput-oriented requests for back-of-house processing (e.g., documents batch processing, referred to best-effort or BE requests), with complex hybrid inference workloads to the underlying model. State-of-the-art (SOTA) LLM serving systems dedicate machines to each type of requests, towards either low inference latency or high serving throughput, respectively. This simplifies request scheduling and management but suffers from poor resource utilization. We propose BROS, a hybrid LLM serving system that aims to collocate RT/BE requests, meeting RT requests’ latency requirements while maintaining BE requests’ throughput. BROS formulates the problem of hybrid RT/BE request scheduling and solves it with a dynamic priority-based algorithm. BROS designs a bidirectional KV cache management mechanism, allowing RT requests to share KV memory with BE requests to remove the scheduling restrictions caused by memory insufficiency and improve utilization. Extensive experiments validate that BROS achieves a good trade-off when serving hybrid RT and BE requests. It significantly reduces the latency of RT requests (up to 74.20%), improving their fine-grained service level objectives (SLOs) attainments (up to 36.38×), with negligible throughput reduction for BE requests, showing significant advantages over SOTA systems.

Federated Learning (FL) enables collaborative model training across distributed devices while safeguarding data and user privacy. However, FL remains susceptible to privacy threats that can compromise data via direct means. That said, indirectly compromising the confidentiality of the FL model architecture (e.g., a convolutional neural network (CNN) or a recurrent neural network (RNN)) on a client device by an outsider remains unexplored. If leaked, this information can enable next-level attacks tailored to the architecture. This paper proposes a novel side-channel fingerprinting attack, leveraging flow-level and packet-level statistics of encrypted wireless traffic from an FL client to infer its deep learning model architecture. We name it FLARE, a fingerprinting framework based on FL Architecture REconnaissance. Evaluation across various CNN and RNN variants-including pre-trained and custom models trained over IEEE 802.11 Wi-Fi-shows that FLARE achieves over 98% F1-score in closed-world and up to 91% in open-world scenarios. These results reveal that CNN and RNN models leak distinguishable traffic patterns, enabling architecture fingerprinting even under realistic FL settings with hardware, software, and data heterogeneity. To our knowledge, this is the first work to fingerprint FL model architectures by sniffing encrypted wireless traffic, exposing a critical side-channel vulnerability in current FL systems.

In many communication networks, the transmission of one packet might automatically trigger the transmission of a subsequent packet from the same source after some delay, without requiring acknowledgment or feedback on the status of the primary transmission. Traditional queueing models, assuming independent interarrival times, fail to model such a scenario. In this paper, we analyze a single-server queue with two classes: primary customers arriving via a Poisson process, and secondary customers triggered by primary ones after a random delay. Both have i.i.d. exponential service times. We show that secondary customers see the queue either in steady state or with one extra customer. Based on this, we propose an analytical approach that models secondary arrivals as an independent Poisson process with an effective service time, consisting of an exponential distribution (probability p) and an Erlang-2 distribution (probability 1− p). Consequently, the primary arrival process is thinned by p. The system is modeled as a multi-class M/G/1 queue, enabling the computation of p and the Laplace-Stieltjes transform of the system time. Simulations confirm the high accuracy of our model, particularly for average system time.

Due to the distributed training framework, FL is vulnerable to backdoor attacks from malicious clients. However existing backdoor attack methods have insufficient attack success rates at low attack frequencies, and the attack effects are difficult to sustain. Therefore, we propose a novel backdoor attack method in FL, named Datura. Datura firstly utilizes adversarial samples to obtain the global prediction model. It then leverages the softmax probability vector to design a trigger generation loss to optimize pattern and mask, creating a highly adaptable trigger. This stage makes it possible to achieve a high attack success rate with minimal attack frequencies. Secondly, in order to obtain the durable high attack success rate, Datura leverages the decentralization of model updates and introduces a layered poisoning strategy based on the redundancy of the model parameters, applying ‘heavy poisoning’ to parameters with high redundancy and ‘light poisoning’ to the rest. We compare Datura with five representative backdoor attack methods on six datasets. Detailed experimental results demonstrate that Datura achieves an attack success rate exceeding 95% with just 1 - 2 attack frequencies, defeating the six defense methods. The 90%-Lifespan of backdoors implanted by Datura reaches an average of 845 rounds after the attack stops.

Owing to its quick reaction to link failures, fast reroute is among the most popular survivable routing approaches in carrier-grade backbone networks. By using preconfigured routing tables, routers can select a failover path for packets based solely on locally available information. Although arc-disjoint spanning arborescences are frequently used to configure forwarding tables in fast reroute, they only provide survivability up to the global connectivity of the network, thus, they fail to maximize survivability in topologies with dense subgraphs. In this paper, we investigate routing arborescences, a generalization of spanning arborescences for fast reroute in which the arborescences are not required to span the entire network. We prove a surprising graph-theoretical result: for any chosen root node in real-world topologies with bidirectional communication links perfect routing arborescences always exist, i.e., each node with local connectivity k to the root is included in exactly k arc-disjoint arborescences. Our constructive proof gives a fast heuristic algorithm to build perfect routing arborescences, offering four optimization options to minimize path length as its primary objective. Extensive simulations are conducted on real-world topologies, demonstrating that our method significantly improves path stretch in the routing arborescences compared to previous methods.

Ethereum 2.0 adopts a new publish/subscribe-based network protocol stack to support its peer-to-peer (P2P) architecture for exchanging consensus messages under the Proof-of-Stake mechanism. Each node scores its peers based on historical behavior and maintains a mesh set of high-scoring peers that can receive and forward full messages. As a result, message propagation time and bandwidth consumption are highly sensitive to the mesh set size. Despite its critical role, optimizing the mesh size to balance propagation delay and bandwidth cost has received limited attention. To fill this gap, we propose the Ethereum 2.0 Network Analyzer (Eth2.0-NA), a framework that optimizes mesh size through two key contributions: (1) a theoretical model that quantifies the relationship between message propagation hop count and mesh size, enabling mesh size optimization for a given hop count target; and (2) a distributed measurement architecture employing multiple probe nodes to empirically validate the model. Experimental results confirm that the theoretical model aligns closely with real-world network behavior. Leveraging this model, we recommend reducing the default mesh size from 8 to 7 on the Ethereum 2.0 MainNet, achieving a 14.1% reduction in bandwidth consumption without increasing propagation delay.

The proliferation of Artificial Intelligence (AI) applications on edge devices requires frequent edge-assisted model retraining to mitigate data drift. This creates significant resource contention on edge servers, exacerbated by pre-committed high-priority reserved tasks. Instead of investing in costly dedicated servers, we identify a key and overlooked opportunity lying in Underutilized Edge Computing (UEC) resources, which arise from the over-provisioning and fragmentation of these reserved tasks. However, existing scheduling paradigms, designed for dedicated resources, are ineffective at leveraging transient and fragmented UEC environment. To bridge this gap, we introduce RESCUE, a novel online framework that unlocks the UEC’s value for opportunistic retraining. RESCUE enables real-time decision-making under uncertainties in task arrival and retraining duration. It facilitates heterogeneous resource-task assignments and co-schedules reserved tasks to maximize expected profit. We formulate this challenge as a stochastic joint-optimization problem and propose a two-stage approach: the offline stage establishes an optimal benchmark and value functions, while the online stage uses this guidance to make irrevocable scheduling decisions in an uncertain environment. We prove RESCUE achieves a competitive ratio of 1/2. Extensive simulations consistently yield a high empirical competitive ratio between 0.51-0.87, and an average profit increase of over 60% compared to baseline methods. These results indicate RESCUE’s robustness and scalability, establishing its effectiveness for real-world edge systems.

Recent advancements in large language models (LLMs) have catalyzed a substantial surge in demand for LLM services. While traditional cloud-based LLM services satisfy high-accuracy requirements, they fall short in meeting critical demands for low delay and enhanced privacy. To address these limitations, we propose HAT, a novel device-cloud collaborative inference framework that leverages the complementary strengths of U-shaped inference and speculative decoding. HAT partitions the LLM into three submodels, and the input and output submodels, stacked with a lightweight adapter network, are deployed as a small language model (SLM) on each end device. Meanwhile, the middle submodel, encompassing the majority of the LLM’s decoder layers, is hosted in the cloud to perform speculative decoding with on-device SLMs. During inference, HAT exchanges hidden states (rather than raw tokens) of input or draft tokens between devices and the cloud, thereby incurring substantial communication delays. Besides, processing hidden states of long prompts will exacerbate computation delays in the cloud, further compromising inference efficiency. To improve efficiency, we introduce a chunking pipeline mechanism that segments long prompts into shorter chunks, enabling parallel transmission and processing. Furthermore, HAT is implemented to dynamically determine optimal chunk sizes for devices handling long prompts, thereby improving overall inference speed. Extensive experiments are conducted on a physical testbed comprising 30 NVIDIA Jetson devices and a server with 8 NVIDIA A6000 GPUs. Experimental results demonstrate that HAT achieves promising performance improvements, reducing TTFT by 41% to 54% and TBT by 41% to 77% compared to the baselines.

Emerging drone delivery systems face challenges in last-meter package deployment due to positioning inaccuracies of GPS. Ultra-wideband (UWB) technology offers high-precision ranging, and when combined with IMU-based dead reckoning, it can enhance drone navigation to enable precise arrival at the destination. However, drone-to-tag UWB localization presents several challenges, including flip ambiguity, IMU drift, and horizontal insensitivity due to high-altitude flights. To address the challenges, we propose NOVA, a navigation optimization scheme via UWB-assisted iterative path planning. With NOVA, the drone follows a trajectory that mitigates flip ambiguity while an iterative joint optimization process simultaneously estimates the UWB tag’s position and refines the drone’s trajectory. At the vicinity of the tag, the drone performs subtle 2D corrective maneuvers during descent to ensure precise last-meter positioning. We evaluate NOVA via real experiments and simulations, demonstrating an average navigation error of 0.83 m with 25.45% trajectory overhead.

Cloud-based autonomous driving relies on low-latency video transmission for accurate downstream perception tasks. However, low-latency video streaming solutions, optimized for human Quality of Experience (QoE), are misaligned with the needs of machine vision tasks. While Region-of-Interest (ROI) encoding aims to bridge this gap, existing methods suffer from two critical limitations: first, insufficient protection for small and vulnerable objects, which are often missed by detection mechanisms or given uniform, inadequate resource allocation; second, prohibitive encoding overhead from fine-grained partitioning, leading to poor system robustness in dynamic network environments. To address these issues, we propose SpotStream, a novel low-latency video transmission framework. SpotStream features a dual-stream importance prediction network to ensure comprehensive coverage and prioritized protection for small objects. It couples this with an efficient, multi-level ROI assignment strategy that significantly reduces encoding overhead by simplifying region structure while focusing resources on critical targets. Comprehensive evaluations demonstrate that SpotStream substantially outperforms state-of-the-art baselines. Notably, under challenging real-world network traces, Spot-Stream improves the F1 score for critical small objects by an average of 12.9% while maintaining excellent system robustness.

Batching is a critical technique for improving energy efficiency (EE) when dealing with edge-based inference tasks. While existing batching policies achieve significant energy savings by exploiting short-timescale arrival patterns, they largely overlook long-timescale load fluctuations—leading to severe EE degradation during lightly-loaded periods, as confirmed by prototype measurements. To overcome this limitation, we propose Long-Short-Timescale Cooperative Batching (LSTCB), a novel framework that enhances EE without sacrificing latency. LSTCB dynamically deactivates servers in response to load variations while intelligently coordinating batched inference tasks. To achieve this, we first develop a measurement-based EE model for edge servers that accurately characterizes the relationship among batch size, workload, and energy consumption. At the short timescale, we devise a dynamic batching policy through Lyapunov optimization that simultaneously minimizes both energy consumption and latency under varying traffic conditions. For long-timescale optimization, we implement an intelligent server deactivation scheme that employs dynamic programming combined with bipartite graph matching to steer inter-server traffic, thereby minimizing overall system-level energy costs while maintaining service quality. Real-world prototype experiments demonstrate that LSTCB achieves 24.48%-46.12% energy savings over baseline methods while maintaining competitive latency performance.

Curving beams are a promising new method for bypassing obstacles in future millimeter-wave to sub-terahertz (sub-THz) networks but lack a general predictive model for their reflections from arbitrary surfaces. We show that, unfortunately, attempting to "mirror" the incident beam trajectory across the normal of the reflector, as in ray optics, fails in general. Thus, we introduce the first geometric framework capable of modeling the reflections of arbitrary convex sub-THz curving beams from general reflectors with experimental verification. Rather than "mirroring" the trajectory, we decompose the beam into a family of tangents and demonstrate that this process is equivalent to the Legendre transform. This approach allows us to accurately account for reflectors of any shape, size, and position while preserving the underlying physics of wave propagation. Our model is validated through finite element method simulations and over-the-air experiments, demonstrating millimeter-scale accuracy in predicting reflections. Our model provides a foundation for future curving beam communication and sensing systems, enabling the design of reflected curved links and curving radar paths.

Emerging Low Earth Orbit (LEO) satellite constellations are revolutionizing global connectivity. However, deploying independent constellations for every operator is infeasible due to limited spectrum and orbital resources, leading operators (e.g., mobile network operators) to lease satellites from satellite network operators (SNOs). This outsourcing model deprives leasing operators of control over service quality, as SNOs may underdeliver without effective enforcement. We advocate: Don’t police the satellite, let the market police it. We present Ripple, turning competition into a built-in quality enforcer. Ripple empowers users to pay for desired QoS through postpaid plans or tokens, measure their runtime QoS loss induced by SNOs, and penalize underperforming SNOs through revenue loss by switching to their competitors. Ripple further empowers leasing operators to stimulate competition among SNOs via performance-adaptive agreements, incentivizing SNOs to maintain optimal service. Our game-theoretical analysis proves that the optimal strategy for SNOs to maximize their revenues is to honestly deliver the best service they can offer, enabling end-users and operators to indirectly regain control of service quality. We implement Ripple in direct-to-cell satellite networks and conduct real trace-driven large-scale simulations. Results show Ripple ensures users receive optimal service while awarding 27.7% higher revenue to SNOs consistently delivering optimal service.

Single-layer networks, whether ground base stations (GBSs) or aerial base stations (ABSs), struggle to provide mobile aerial user equipment (aUEs) with ultra-reliable low-latency communication (URLLC) and multi-access edge computing (MEC) services. These inherent limitations—coverage gaps, computational bottlenecks, and communication delays—are exacerbated by UAV mobility, necessitating effective cooperative handoff management in integrated air-ground heterogeneous networks (AGHetNets) to ensure service continuity and stringent QoS. To address this, we propose a novel Poisson-Delaunay tetrahedralization (PDT)-based AGHetNet architecture. This network employs four-node Coordinated Multi-Point (CoMP) sets (combining ABSs and GBSs) to provide URLLC and MEC for aUEs. Leveraging stochastic geometry, we design a cooperative handoff strategy for arbitrary aUE trajectories and derive an analytical expression for CoMP handoff probability. Furthermore, we propose a generalized framework for the aggregate interference power distribution incorporating NLoS probability, followed by characterizing the successful uplink communication probability and the successful edge computing probability based on handoff dynamics. Simulations demonstrate our model’s significant superiority over traditional Voronoi approaches, reducing handoff probability by 28% on average and enhancing both CoMP transmission and MEC performance. Crucially, our PPP-based handoff analysis provides a robust approximation for practical deployments, offering vital design parameters for future AGHetNets.