INFOCOM 2026 Papers

Congestion Control Algorithm (CCA) and paced sender play a crucial role in real-time communication (RTC) applications. While CCA has been studied extensively, the paced sender remains largely unexplored. The default paced sender follows a fixed-interval packet-sending schedule, and employs a uniform budget distribution approach, which is not suitable for RTC applications. With a thorough analysis of the default paced sender, we reveal that the end-to-end delay and CPU cost are far from satisfactory in large-scale deployments.In this paper, we propose SmartPacer, a novel approach that ensures both low latency and high efficiency for RTC applications, even in dynamic and fluctuating network environments. SmartPacer is designed with two key components: a hierarchical pacer budget control scheme and an adaptive scheduling strategy. By integrating SmartPacer with some state-of-the-art bandwidth estimation algorithms, we show that SmartPacer helps achieve a 10%–50% reduction for end-to-end delay. Furthermore, we deploy SmartPacer on Douyin, one of the most popular video platforms in China, where 120 million users are involved in the online A/B test. Comprehensive results demonstrate that CPU usage is reduced by 14.4%, and end-to-end delay is reduced by 7.6%. More importantly, the user engagement duration is increased by 1.0%.

The Open Radio Access Network (ORAN) architecture promotes flexibility and innovation by allowing independently developed xApps and rApps to control network behavior. However, this disaggregated design also introduces risks of resource conflicts that are difficult to detect and resolve, especially in unseen or dynamic environments. We propose ORCA, a conflict management framework that leverages generative LLM agents to reason about the importance of control parameters under environmental uncertainty. The framework incorporates symbolic reasoning through a novel β-Reasoning Cascade and introduces Segmented Active RAG (SA-RAG) to retrieve relevant domain knowledge dynamically. Evaluations on diverse environments show that ORCA reduces cosine error by 60.7% and NRMSE by 65.1% over the best baselines, achieving the most accurate feature importance estimates for conflict detection. Our findings demonstrate the potential of LLM-driven reasoning to manage conflicts in complex and dynamic RAN systems.

Service caching facilitates service provisioning in mobile edge computing (MEC) systems, which support latency-sensitive and computationally-intensive applications (services) such as intelligent traffic systems and online video games. In such systems, users are served by service data cached locally at an edge server (ES). The ES may use its cached service data to serve requests from its users and allocate its computation resources to the users, or offload the requests to the cloud center (CC). The ES should also download refreshed versions of the service data from the CC to maintain the freshness of its cached service data. In this paper, we investigate the joint online design of fresh service caching, task offloading, and computation resource allocation of the ES. The formulated problem involves spatially and temporally coupled variables, which makes the problem challenging to handle. To solve the formulated problem, we propose a two-layer online algorithm termed Knapsack Problem-based Double Proximal policy optimization (KPDP) algorithm, in which the first layer addresses service caching by using proximal policy optimization (PPO) and knapsack problem formulation, while the second layer addresses task offloading and computation resource allocation by using another PPO. We also perform a competitive ratio analysis for our proposed algorithm. Through extensive simulation experiments, it shows that KPDP algorithm outperforms other benchmark algorithms.

In this paper, we propose Adriana, the first provable resource-aware decentralized learning framework that leverages rate-adaptive quantization to enhance both computational and communication efficiency. Theoretically, we present a rigorous analysis of the proposed Adriana algorithm and formally characterize how training and communication quantization errors, caused by heterogeneous and limited client resources, jointly affect the convergence of decentralized learning. When these errors are negligible, Adriana converges to a small constant with a rate of $\mathcal{O}\left( {\frac{1}{{\sqrt {nRT} }}} \right)$, which matches the state of the arts under resource-unconstrained scenarios. Then we formulate the quantization scale selection problems and propose rate-adaptive one-block and multi-block communication quantization schemes that enable efficient solutions, thereby accelerating convergence. Extensive experiments on diverse datasets demonstrate that Adriana reduces local computation by 62.63%–71.70% and communication costs by approximately 50%, without sacrificing accuracy compared to full-precision DPSGD.

Ultra-low-power (ULP) Wi-Fi (consuming <100µW) is critical for creating battery-free IoT devices that operate via energy harvesting. While backscatter techniques effectively address the uplink, demodulating complex downlink Wi-Fi signals remains a major bottleneck due to the severe power constraints of ULP nodes. Existing downlink solutions are either slow (around 125 kbps) or require custom modifications to commercial off-the-shelf (COTS) Wi-Fi devices. In this paper, we present LiteWiFi, a system that achieves high-speed Wi-Fi demodulation at the microwatt level without requiring hardware modifications to COTS Wi-Fi infrastructure. LiteWiFi introduces a passive mixing model for OFDM signals and a lightweight passive alignment technique for eliminating phase errors. LiteWiFi achieves a data rate of up to 2.22 Mbps, representing a 17.8X improvement over prior symbol-level systems and a 1.97X improvement over state-of-the-art custom hardware solutions.

Integrated sensing and communication (ISAC) has emerged as a key technology for improving spectrum efficiency while delivering dual functionalities. Recent work has integrated FMCW radar sensing into modern OFDM-based wireless systems to enable fine-grained, low-cost sensing with orthogonal coexistence. However, these approaches rely on fixed spectrum allocation, reserving static pilot subcarriers for sensing, which makes them vulnerable to frequency-selective fading and channel fluctuations. To overcome this limitation, this paper presents DyPilot, a Dynamic Pilot configuration framework. The core idea is a novel FMCW waveform design that explicitly controls pilot placement, combined with a learning protocol that adaptively adjusts this placement in response to changing channel conditions. Evaluations show that DyPilot consistently achieves higher data rates and lower sensing errors than random and sampling-based schemes across diverse dynamic environments.

Backscatter is promising to deliver near-zero-power communications for billions of IoT devices. While significant progress has been made in uplink transmission, existing systems suffer from impractical downlink design, failing to interpret packets from ambient wireless networks. This prevents tags from applications that necessitate two-way communication, such as resource allocation or reliable backscatter services. We propose InterfereX, a novel system that enables practical downlink communication for backscatter tags with commodity radios. The key behind InterfereX is an interference-driven demodulation scheme that transforms phase shifts of ambient RF signals into amplitude-modulated waveforms. By leveraging differential phase modulation inherent in protocols like ZigBee, Bluetooth, and 802.11b WiFi, InterfereX enables energy-efficient decoding through envelope detection. We prototype InterfereX and evaluate its performance through extensive experiments. Results show that InterfereX achieves downlink throughputs of 249.3 kbps for ZigBee, 994.8 kbps for Bluetooth, and 1626.3 kbps for WiFi, with BERs as low as 0.39‰, 5.98‰, and 4.90‰. The sensitivity of ZigBee/Bluetooth/802.11b WiFi demodulation reaches -35.5 dBm, -32.5 dBm, and -30.5 dBm. These results confirm that InterfereX provides a practical downlink, enabling seamless integration of backscatter tags into wireless infrastructure.

Virtual Private Networks (VPNs) are widely used for censorship evasion and traffic protection. VPN users expect to be provided with adequate security protection, and at the same time not be affected by other users connected to the same VPN server, which can be illustrated as the non-interference property. However, in this paper, we have identified several vulnerabilities that violate this property, specifically within the connection tracking frameworks of VPN servers, stemming from shared resource misuse and insufficient validation of session state transitions. We present three session manipulation attacks targeting TCP and UDP traffic tunneled through VPNs. The attacker who only connects to the same VPN server can launch denial-of-service attacks, hijack TCP connections of other clients, or inject forged DNS responses into their queries. We evaluate these attacks against five popular connection tracking frameworks across different OSes and nine major commercial VPN providers. Experimental results reveal that all frameworks and eight providers are vulnerable to at least one of the attacks. We have responsibly disclosed our findings with countermeasures, resulting in 19 assigned CVEs/CNVDs and acknowledgments from the communities and providers.

The Traffic Matrix (TM) plays a vital role in a wide range of network management tasks, yet monitoring complete TMs is prohibitively expensive. Recent efforts focus on recovering incomplete TMs via learning-based models by learning from historical TM snapshots. However, existing approaches often suffer from poor generalization and require frequent retraining when the network environment evolves—let alone unseen networks. In this paper, we propose a novel TM imputation framework, named ARI-LLM, based on Large Language Models (LLMs). Inspired by human-like cognitive process of coarse-to-fine reasoning, our method incrementally refines the incomplete TM through a step-wise autoregressive mechanism, which is compatible with the pretraining paradigm of LLMs, endowing our model with powerful generalization and zero-shot capabilities. To ensure plug-and-play deployment across TM imputation tasks in dynamic environments, we craft a flow-aware tokenization scheme that supports scalable and topology-adaptive embeddings. Extensive experiments demonstrate that ARI-LLM significantly outperforms state-of-the-art methods on large-scale sparse TMs (over 90% missingness), reducing imputation error by up to 34.7%. It maintains strong performance whether given only limited TM knowledge (e.g., 20% observed flows, 5% training snapshots) or no prior knowledge at all on unseen networks.

Free-viewpoint video (FVV) enables immersive viewing experiences by allowing users to view scenes from arbitrary perspectives. As a prominent reconstruction technique for FVV generation, 4D Gaussian Splatting (4DGS) models dynamic scenes with time-varying 3D Gaussian ellipsoids and achieves high-quality rendering via fast rasterization. However, existing 4DGS approaches suffer from quality degradation over long sequences and impose substantial bandwidth and storage overhead, limiting their applicability in real-time and wide-scale deployments. Therefore, we present AirGS, a streaming-optimized 4DGS framework that rearchitects the training and delivery pipeline to enable high-quality, low-latency FVV experiences. AirGS converts Gaussian video streams into multi-channel 2D formats and intelligently identifies keyframes to enhance frame reconstruction quality. It further combines temporal coherence with inflation loss to reduce training time and representation size. To support communication-efficient transmission, AirGS models 4DGS delivery as an integer linear programming problem and design a lightweight pruning level selection algorithm to adaptively prune the Gaussian updates to be transmitted, balancing reconstruction quality and bandwidth consumption. Extensive experiments demonstrate that AirGS reduces quality deviation in PSNR by more than 20% when scene changes, maintains frame-level PSNR consistently above 30, accelerates training by 6×, reduces per-frame transmission size by nearly 50% compared to the SOTA 4DGS approaches.

Federated learning (FL) enables collaborative model training without sharing raw data, but also raises increasing demands for the right to be forgotten. To support data erasure, active federated unlearning (FU) allows clients to actively remove their data’s influence from the model. We reveal a critical and overlooked threat: malicious clients can pose as privacy-concerned users requesting to unlearn some of their data, while secretly preparing backdoor attacks during training. We propose FUsion backdoor, a subnetwork-based attack that stealthily constructs a compact backdoor subnetwork from trigger-sensitive units within backdoor-critical layers during training, and rapidly fuses it during the limited rounds of unlearning. FUsion backdoor achieves up to 99% backdoor success rate across diverse datasets and FU methods. We also develop a detection method that captures directional subspace deviations introduced by coordinated backdoor updates, achieving high attack detection accuracy.

The rapid proliferation of network devices has created a critical security challenge: widespread deployment of systems running vulnerable firmware versions. Firmware version identification is essential for security management, but existing methods face a fundamental limitation: they cannot simultaneously achieve fine-grained accuracy and large-scale practicality. In this paper, we present Chronos, a novel approach exploiting temporal signatures inadvertently embedded during firmware development. This enables fine-grained version identification without firmware access, achieving both accuracy and large-scale practicality. We evaluate Chronos on 7,740 real-world samples across 9 brands, achieving 84% fingerprint generation coverage with 96.5% precision and 91.8% recall. Our large-scale measurement reveals widespread security vulnerabilities: only 14.4% of devices run current firmware, with 17.2% exposed to known vulnerabilities—69.1% classified as High or Critical severity. The code is available at https://doi.org/10.5281/zenodo.18175223.

Eavesdropping on voice conversations presents a growing threat to personal privacy and information security. In this paper, we present RadEar, a novel RF backscatter-based system designed to enable covert voice eavesdropping through walls. RadEar consists of two key components: (i) a batteryless RF backscatter tag covertly deployed inside the target space, and (ii) an RF reader located outside the room that performs signal demodulation, voice separation, and denoising. The tag features a compact, dual-resonator design that achieves energy-efficient frequency modulation for continuous voice eavesdropping while mitigating self-interference by separating excitation and reflection frequencies. To overcome the challenges of weak signal reception and overlapping speech, the RF reader employs self-supervised learning models for voice separation and denoising, trained using a remix-based objective without requiring ground-truth labels. We fabricate and evaluate RadEar in real-world scenarios, demonstrating its ability to recover and separate human speech with high fidelity under practical constraints.

Radio Access Networks (RANs) are critical infrastructures that mobile operators continuously upgrade to accommodate increasing data traffic demands, stricter performance requirements, and evolutions in radio technologies. RAN updates can affect carrier-level Key Performance Indicators (KPIs) that are the foundational input to data-driven models for network management. However, to date, no study has systematically examined the dynamics of RAN deployments, and little is known about the actual prevalence of RAN updates or their impact on Machine Learning (ML) models for network automation. This paper presents a first characterization of RAN updates in a nationwide operational infrastructure composed of over 500,000 carriers. A network-side vantage point lets us (i) investigate the type and frequency of RAN modifications, (ii) assess the impact of such changes on a primary KPI for network management, i.e., the traffic volume served by individual carriers, and (iii) verify the final effects on a classical downstream ML application, i.e., traffic prediction. Our results reveal that RAN updates take place with notable frequency, e.g., occurring every few days even in medium-sized cities. Also, they affect in a significant way the demands at a considerable fraction of pre-existing carriers, where they can curb the accuracy of ML traffic forecasting models.

The rollout of 5G represents a significant advancement in the telecommunications industry, offering the potential for markedly enhanced speeds, reduced latency, and improved connectivity. Considering these anticipated advantages, it is interesting to understand the progressive adoption of the new technology by operators and their subscribers. In this paper, we analyze the evolution and current operation of the nation-wide 5G network of Orange, a leading mobile operator in France. By inspecting longitudinal data about (i) the over-five-year-long development of the country-wide 5G radio access infrastructure and (ii) the last two years of 5G traffic demands, we unveil how the operator has planned the deployment of the 5G radio access and characterize the actual usage patterns of the available 5G infrastructure. We also investigate the recent introduction of a 5G Standalone (SA) commercial service and its adoption by the mobile subscribers. We show that by mid 2025, the 5G network under study has achieved substantial coverage of populated areas and the operator has very recently started adding capacity layers to its 5G radio access. However, our investigation reveals that such massive infrastructure deployment efforts are not matched by a commensurate adoption of the technology by the end users, as the 5G capacity -especially for SA- stays largely underutilized.

The sparsely activated mixture-of-experts (MoE) transformer has become a common architecture for large language models (LLMs) due to its sparsity, which requires fewer computational demands while easily scaling the model size. In MoE models, each MoE layer requires to dynamically choose tokens to activate particular experts for computation while the activated experts may not be located in the same device or GPU as the token. However, this leads to substantial communication and load imbalances across all GPUs, which obstructs the scalability of distributed systems within a GPU cluster. To this end, we introduce HierMoE to accelerate the training of MoE models by two topology-aware techniques: 1) token deduplication to reduce the communication traffic, and 2) expert swap to balance the workloads among all GPUs. To enable the above two proposed approaches to be more general, we build theoretical models aimed at achieving the best token duplication and expert swap strategy under different model configurations and hardware environments. We implement our prototype HierMoE system atop Megatron-LM and conduct experiments on a 32-GPU cluster with DeepSeek-V3 and Qwen3-30B-A3B models. Experimental results show that our HierMoE achieves 1.55× to 3.32× faster communication and delivers 1.18× to 1.27× faster end-to-end training compared to state-of-the-art MoE training systems, Tutel-2DH, SmartMoE, and Megatron-LM.

Intent-Based Networking (IBN) has seen rapid advancement in recent years, largely fueled by the capabilities of large language models (LLMs). While LLMs enable more expressive and automated intent translation solutions, challenges such as hallucinated parameters, missing constraints, and syntax errors still limit their practical deployment. To address these challenges, we propose Artanis, an intent-based configuration framework that combines LLM-powered intent translation with multi-objective verification feedback to iteratively update configurations. Artanis comprises a Translator that produces candidate configuration updates (CUs) from intents, and a scalable multi-objective Verifier to evaluate the CUs. To enhance service guarantee, we introduce Group Relative Policy Optimization (GRPO) with verification feedback, a reinforcement learning method that dynamically adjusts the CUs by leveraging structured feedback from the Verifier. Evaluation on comprehensive network topologies shows that Artanis reduces the policy violation rate by 32.6% compared to existing approaches. Artanis achieves this while maintaining consistency with intents and adapting robustly to diverse network conditions.

Online learning is a foundational machine learning paradigm in both academia and industry. Most existing online learning techniques are designed for idealized scenarios where the feedback on the decision costs observed by the learner is assumed reliable, i.e., the same as the ground truth. Many recent efforts that attempted to investigate erroneous feedbacks considered errors with unrealistic settings, such as restrictive stochasticity and/or error bounds (e.g., bounded corruption magnitudes or budget of deviations from the ground truths). In this work, we consider a novel and challenging problem of full-information online learning in the presence of feedback with non-stochastic and unbounded errors. According to our analysis, existing representative techniques suffer unbounded regret when applied to our problem. To tackle such erroneous feedback, we propose a robust online learning strategy with a tailored FTRL-like decision-making approach based on a coordinate-wise trimmed sum mechanism, which we prove can achieve a near-optimal, sublinear regret bound of ${\mathcal{O}}\left({\sqrt T }\right)$ under certain justified conditions. We compare our solution against four representative approaches by evaluating them in three exemplary networking applications. The results not only corroborate our theoretical analysis but also clearly demonstrate the robustness of our algorithm in comparison to the baselines.

In-vehicle wireless systems, such as vehicle-to-satellite communication, face challenges in adapting to rapidly changing environments due to the inherent latency of aligning high-precision beamforming with distant base stations (e.g., satellites). Motivated by the robust sensing capabilities of today’s autonomous vehicles, we propose modeling the radio environment in parallel. This allows the in-vehicle communication subsystem to predict channel conditions along the vehicle’s planned trajectory and precompute optimal beamforming parameters in advance, effectively addressing rapid environmental changes. To this end, we introduce Full Smart-Communication (FSC), a novel framework that repurposes intermediate feature representations—such as planned trajectories, occupancy maps, and posture estimates—from a vehicle’s end-to-end Full Self-Driving (FSD) model to derive spatio-temporal stochastic radio radiance fields. This approach avoids direct access to raw sensor streams while effectively correlating kinematic and environmental states with radio propagation dynamics. Our comprehensive experiments demonstrate that FSC achieves a mean gain of 15 dB and maintains less than one-degree beamforming error at direct upcoming positions. Additionally, FSC reduces power consumption by 70% and GPU memory usage by 94% through the hitchhiking approach. By leveraging FSD-derived environmental awareness for communication, our work bridges the gap between autonomous navigation and robust connectivity, unlocking the synergy of perception and radio resource management in next-generation self-driving vehicles.