INFOCOM 2026 Papers

Multi-participant meetings occur across various domains, such as business negotiations and medical consultations, during which sensitive information like trade secrets, business strategies, and patient conditions is often discussed. Previous research has demonstrated that attackers with mmWave radars outside the room can overhear meeting content by detecting minute speech-induced vibrations on objects. However, these eavesdropping attacks cannot differentiate which speech content comes from which person in a multi-participant meeting, leading to potential misunderstandings and poor decision-making. In this paper, we answer the question ``who speaks what''. By leveraging the spatial diversity introduced by ubiquitous objects, we propose an attack system that enables attackers to remotely eavesdrop on in-person conversations without requiring prior knowledge, such as identities, the number of participants, or seating arrangements. Since participants in in-person meetings are typically seated at different locations, their speech induces distinct vibration patterns on nearby objects. To exploit this, we design a noise-robust unsupervised approach for distinguishing participants by detecting speech-induced vibration differences in the frequency domain. Meanwhile, a deep learning-based framework is explored to combine signals from objects for speech quality enhancement. We validate the proof-of-concept attack on speech classification and signal enhancement through extensive experiments. The experimental results show that our attack can achieve the speech classification accuracy of up to $0.99$ with several participants in a meeting room. Meanwhile, our attack demonstrates consistent speech quality enhancement across all real-world scenarios, including different distances between the radar and the objects.

Cardiovascular diseases typically require continuous, long-term electrocardiogram (ECG) monitoring to detect intermittent cardiac abnormalities. Currently, ECG monitoring generally relies on wearable devices, which face the issues of limited battery life and electrode shedding, resulting in monitoring discontinuity. In contrast, mmWave radar enables non-contact vital sign monitoring by detecting subtle chest vibration. Based on this, we present a novel mmWave radar-based ECG monitoring system, RaECG, which perceives both heartbeat-induced chest vibration and carotid pulse from the neck and employs spatiotemporal feature fusion to improve ECG reconstruction accuracy. RaECG adopts a respiration-guided interference suppression by enhancing the amplitude of the echo signals caused by chest vibrations and a dynamic time warping (DTW)-based signal evaluation algorithm to accurately capture the signals from heartbeat-induced chest vibration and carotid pulse. Then, a parallel neural network is designed to extract spatiotemporal features, where the Squeeze-and-Excitation-enhanced TCN module dynamically weights features from the chest vibration and carotid pulse, combined with the Global Attention-enhanced BiGRU to model ECG periodicity. The experimental results show the effectiveness of RaECG in ECG reconstruction and achieve high morphology accuracy with a median Pearson correlation coefficient of 0.941 and a median root mean square error of 0.078mV in amplitude.

Symmetric key cryptography is commonly employed in wireless networks, and is considered to be secure even in the emerging quantum computation paradigm. Typically, the establishment of a shared secret key is achieved through asymmetric cryptography, in which legitimate users leverage computationally acquired public and private key pairs to agree on a symmetric key. Unfortunately, this method requires a trusted third party, introduces computational costs to users, and incurs delays on the order of 10s to 100s of milliseconds for Internet round-trip time to connect with the third party. Moreover, asymmetric cryptography has well-known vulnerabilities to quantum computing. In this work, we propose a rapid picosecond-scale symmetric key generation method via a free-space wireless channel and without requiring a third party. In particular, we devise a technique for the legitimate users to use transmissive reconfigurable metasurfaces to generate random perturbations in their free-space channel from which they can extract reciprocal random frequency signatures and thus extract random bits. Our method also ensures key confidentiality by spatially randomizing the modulated pilot signal, such that any adversarial node in the environment obtains a completely different bit sequence compared to the legitimate users. We experimentally demonstrate our method and show that the extracted keys pass the standard randomness tests and that our method renders eavesdroppers’ observations statistically independent from those of the legitimate users, provided that they have more than a few degrees of angular separation. Due to our use of picosecond-scale pulses as pilot signals, this method yields several orders of magnitude increase in the bit generation rate compared to prior methods that leverage mobility-induced changes in channel response, while maintaining a low bit mismatch rate.

Digital Twin (DT) technology enables creating high-fidelity virtual models of physical objects, which must be dynamically updated to capture inherent evolutions. Conventional cloud-based schemes often suffer from excessive costs and limited data access, while updating the entire DT model on distributed edge servers (ESs) may cause inconsistency and incompleteness due to prevalent data silos. In response, we propose DTAS, an edge-cloud collaborative DT update framework via adaptive model splitting. In DTAS, the DT model is viewed as a global one composed by multiple elementary units, which are split into disjoint partial-DTs constructed on ESs using locally collected data, and then integrated by the cloud for update. To enhance reliability against uncertain data distortions, after splitting, we allow each partial-DT to be constructed via crowdsourcing, i.e., being built and aggregated by multiple ESs before uploading to the cloud. We formulate an online optimization problem to maximize the physical-virtual mapping accuracy in the dynamic DT update. By leveraging coalitional game, we first derive the solution for a simplified short-term problem, and then extend it adopting deep reinforcement learning to a long-term one, thereby solving the original problem. Simulations show that DTAS is not only effective for the dynamic DT update, but also superior over counterparts with an improvement of physical-virtual mapping accuracy by 25.76% on average.

Over-the-air federated learning (OTA-FL) harnesses the superposition property of wireless channels for efficient model aggregation, offering significant spectrum savings. However, practical deployment is challenged by analog distortions, synchronization errors, and dynamic gradient scaling across neural network layers. This paper presents a robust OTA-FL framework that systematically addresses these limitations. We introduce a gradient-aware power scaling technique that adjusts per-layer transmit power to mitigate the impact of uneven gradient magnitudes, coupled with channel pre-equalization for stable convergence. We implement OTA-FL on a real-world testbed with a multi-antenna gNB and develop a deep learning based multi-antenna combiner for scalable aggregation in dense networks. To enable reproducible and site-aware evaluation, we propose a digital twin framework that integrates ray-traced urban channels into a hardware-in-the-loop testbed via real-time channel emulation. Experimental results on a wireless positioning task demonstrate that our OTA-FL system achieves model accuracy comparable to digital FL while providing a 140× reduction in spectrum usage for 20 participants. This work closes the gap between OTA-FL theory and real-world deployment, highlighting practical solutions for scalable and spectrum-efficient FL.

As blockchain applications become increasingly specialized and fragmented, cross-chain smart contract transactions are essential for platform interoperability. The inherent, significant confirmation delays of these transactions, in serial execution mode, lead to substantial degradation of system performance. To address this issue, existing solutions typically employ Two-Phase Locking (2PL) and Optimistic Concurrency Control (OCC) to support concurrent transaction processing. 2PL-based approaches result in transaction blocking due to continuous resource locking, which substantially degrades system performance. OCC-based approaches frequently encounter transaction aborts and rollbacks due to resource modification conflicts, causing resource wastage. To overcome these challenges, this paper proposes MVCX, a novel concurrency control solution that supports lock-free parallel execution of cross-chain smart contract transactions without sacrificing atomicity. MVCX adopts a speculative execution model driven by multi-version management, combined with conflict detection and re-execution of conflicting transactions to eliminate locking operations. Additionally, MVCX introduces Composite Transaction Units (CTUs) and a depth-first construction (DFC) algorithm, along with a deferred commit mechanism, ensuring that transactions are committed in accordance with their dependency order to maintain correctness and atomicity. Experimental results using SmallBank and YCSB workloads under varying data contention levels demonstrate that MVCX reduces transaction latency by 65% and abort rates by 55% on average, compared to AtomCI and Spectrum.

The increasing demand for private and noise-resilient speech interaction has motivated the development of Silent Speech Interfaces (SSIs) that infer user intent without vocalization. Existing SSI solutions face limitations such as intrusiveness, privacy leakage, environmental sensitivity, deployment complexity, and motion vulnerability. In this work, we present Baro2Talk, a wearable SSI system that reconstructs speech content from TMJ-dominated Pressure Variation Sequences (TPVSs) captured by miniature barometers embedded in standard earbuds. Baro2Talk is inspired by two key observations. First, silent articulation induces consistent ear canal deformation via temporomandibular joint (TMJ) movements, producing pressure fluctuations that reflect articulatory patterns associated with speech. Second, TPVSs exhibit repeatable temporal and articulatory structures within phrases, offering structured signals to support semantic modeling without acoustic input. We develop a lightweight in-ear pressure sensing prototype and propose a set of modules that first perform articulatory event detection and generalization enhancement, followed by a three-stage reconstruction pipeline: Semantic Encoding, Coarse Mel-spectrogram Construction, and Phonetic Enhancement. The resulting spectrograms are decoded into text using a pre-trained automatic speech recognition (ASR) model (e.g., Whisper). Baro2Talk achieves a 6.5% CER, 9.9% WER, and a 0.081 spectral convergence score, demonstrating robust performance in silent, mobile, and noisy environments.

Deploying large multimodal models (LMMs) on the edge is a promising solution for latency-sensitive applications, while also addressing the high bandwidth costs and privacy concerns associated with cloud-based serving. LMM inference involves a compute-intensive context stage and a memory-intensive generation stage. However, existing LMM serving systems typically colocate both stages within a single instance, which precludes independent optimization and leads to inefficient resource utilization and increased latency, especially in resource-constrained edge settings. In this paper, we present DEMUS, a Decoupled edge serving framework for large Multimodal models with diffusion-based Scheduling. DEMUS decouples the two stages to optimize them independently without interference, and schedules context and generation tasks to corresponding instances. It combines diffusion-based generation with model predictive control to effectively schedule tasks across instances on heterogeneous nodes, aiming to maximize overall system throughput under latency constraints. We implement and evaluate DEMUS on a heterogeneous cluster with 16 edge server nodes and 45 GPUs. Extensive experiments shows that DEMUS achieves 2.1 3.7× higher throughput while meeting latency SLOs, and generalizes well in dynamic environments and unseen settings.

Direct-to-cell Low Earth Orbit (LEO) satellite networks have emerged as a transformative solution for global 4G/5G connectivity. However, the unique characteristics of LEO systems, including rapid satellite mobility, long space-to-ground propagation delays, and extensive coverage, challenge conventional paging mechanisms, causing unacceptable latency and signaling overhead. This paper introduces SigHitching, a standards-compliant, non-broadcast paging mechanism that leverages device-to-device (D2D) to relay paging signaling to reach idle User Equipments (UEs) efficiently. Additionally, SigHitching incorporates offline service area configuration and a bitmap-based base station selection algorithm to reduce the paging procedure completion time. We implement a prototype of SigHitching using Open5GS and UERANSIM, and emulate LEO constellations based on real satellite data. Experimental results show that SigHitching achieves a 45.23% and 29.81% improvement in paging latency compared to 5G NTN and satellite-native solutions such as SpaceCore and MOSAIC. These results highlight SigHitching’s potential to improve paging efficiency and enable sustainable, scalable global connectivity.

Today’s massive AI computation loads push heavy data synchronization across sites, i.e., nodes in data centers. Any reduction in such consensus latency can significantly improve the overall performance of desired systems. This consensus challenge explosively peaks at cross-domain sites. In this paper, we proposed CD-Raft to address the cross-domain latency challenge, an optimized Raft protocol for strong consistency in cross-domain sites. CD-Raft can significantly reduce consensus latency by optimizing cross-domain round-trip time (RTT) for reads and writes, as well as carefully positioning the leader node. We verified the correctness of CD-Raft in a formal specification using the TLA+ specification, guaranteeing the strong consistency across sites. We have prototyped CD-Raft and evaluated it using the YCSB benchmark. Empirical results show that compared to the classic Raft, CD-Raft reduces the average latency by 32.90% and (99th percentile) tail latency by 49.24% for renown traces across multiple sites.

Internet-wide service discovery forms the foundation of Internet situational awareness and security research, yet it faces the challenge of identifying services across massive open ports. Existing methods, which rely heavily on extensive prior knowledge (seeds) and suffer from low efficiency, are limited to specific scenarios with small scales and few service types.To this end, we propose SPADE, an efficient Internet-wide service prediction and discovery method with adaptive and divide-and-conquer characteristics. It acquires limited seeds from target hosts through a multi-level adaptive sampling strategy, then progressively extracts two-tier service deployment features and performs multi-layer predictions in phases. This approach significantly enhances prediction performance while substantially boosting efficiency. Across 12 comparative experiments spanning three-level prediction scenarios and four datasets, SPADE achieves over 94% in both Top-1 hit rate and coverage, significantly outperforming the state-of-the-art methods, especially in scenarios where seeds are limited or even scarce. More notably, it only requires 0.15%-1% of the seeds needed by the state-of-the-art approach to surpass its performance. Furthermore, SPADE achieves a 219-805× speedup in service discovery, completing predictions within 7 minutes across all scenarios and datasets (including hundred-million-scale target hosts), while the state-of-the-art method requires ≥ 22 hours.

Wireless Rechargeable Sensor Networks (WRSNs) have emerged as a promising solution to address the energy limitations of wireless sensor networks. However, existing charging optimization methods often assume idealized free-space propagation, neglecting real-world attenuation effects such as variable path loss, shadowing, and multipath fading. This results in overestimated charging efficiency, inaccurate power predictions, and reduced network reliability. In this paper, we focus on optimizing mobile charging and formalize the charging utility maximization problem with attenuation effects (termed OPTIC problem). First, we propose an environment-aware charging model that comprehensively incorporates statistical path loss variations, stochastic shadowing, and multipath fading effects. Then, we construct an attenuation map that partitions the environment into regions with distinct physical characteristics and corrects the theoretical model through statistical analysis of spatial attenuation variations. Additionally, we discretize the charging region and reformulate the OPTIC problem as a constrained submodular maximization problem, which can be solved by a proposed algorithm with an approximation guarantee. Extensive simulations and field experiments demonstrate that our algorithm improves charging utility by 44.3% on average.

As the inter-container data transmission time constitutes a large portion of the overall execution time in contemporary containerized applications, it is natural to group tightly coupled containers into pods to lower the transmission time. However, this container composition leads to longer application waiting time, as all containers within a pod must be started before the pod can provide services. In this paper, we investigate how to group the containers into pods and allocate these pods onto servers with heterogeneous resource capacities and data transmission time, aiming to achieve the minimum completion time. The problem is formulated in quadratic mixed integer programming (QMIP) form and is proven to be NP-hard. Based on the QMIP formulation, we then propose a randomized rounding based multiple containers in one pod (MCOP) placement strategy with reduced computational complexity. Trace-driven experimental results show that MCOP reduces the application completion time by 27.26% to 55.81% compared to state-of-the-art solutions.

With the development of federated learning and mobile computing, federated bandit frameworks have been proposed to enable multiple clients to explore and exploit collaboratively under the guarantee of data privacy. However, none of them take the widespread budget constraints into account. In this work, we introduce the Federated linear Bandits framework with Knapsacks (FBwK), where M clients can pull K arms with linear rewards and costs to minimize the total regret under the coordination of a central server. For FBwK, we propose a new definition of OPT solution and a bandit algorithm, namely FedUCBwK. Specifically, we design a uniform policy parameter update threshold for each client, which balances regret, communication, and computation costs. In this process, the central server uses a factor on knapsack constraints to pace the budget consumption and preserve the clients’ privacy by transmitting model updates instead of raw data. We conduct a theoretical analysis and show that FedUCBwK achieves a sublinear regret with logarithmic communication and computation. Evaluation on real-world mobile application datasets demonstrates that FedUCBwK consistently outperforms existing methods across diverse tasks.

As internet traffic grows, the underlying infrastructure consumes increasing amounts of energy. During off-peak hours, large parts of the networks remain underutilized, presenting significant potential for energy savings. Existing Green Traffic Engineering approaches attempt to leverage this potential by switching off those parts of the networks that are not required for the routing of specific traffic matrices. When traffic changes, the approaches need to adapt rapidly, which is hard to achieve given the complexity of the problem. We take a fundamentally different approach: instead of considering a specific traffic matrix, we rely on a traffic-oblivious routing scheme. We discuss the NP-hard problem of activating as few connections as possible while still guaranteeing that any down-scaled traffic matrix ϱ•T can be routed, where ϱ ∈ (0,1) and T is any traffic matrix routable in the original network. We present a $\max \left( {\frac{1}{{\varrho \cdot {\lambda _{\min }}}},2} \right)$-approximation algorithm for this problem, with λmindenoting the minimum number of connections between any two connected routers. Additionally, we propose two post-processing heuristics to further improve solution quality. Our evaluation shows that we can quickly generate near-optimal solutions. By design, our method avoids the need for frequent reconfigurations and offers a promising direction to achieve practical energy savings in backbone networks.

Regional IP anycast enhances traditional global anycast by organizing infrastructure into geographically-defined regions, each advertising distinct IP prefixes to attract local client traffic, though this approach introduces two key challenges: suboptimal intra-region routing from rigid geographic boundaries and cross-region path inflation due to excessive prefix multi-announcement. To address these, we propose ReOpt, an optimized anycast framework that dynamically minimizes latency through three key mechanisms: (1) real-time RTT measurements between client-site pairs and assess pairwise site preferences for each client, (2) intelligent multi-announcement strategies that enhance routing flexibility while maintaining stability, and (3) country-level region partitioning that simplifies DNS management while preserving geographic optimization. We formulate this as the Latency-Minimized Anycast Region Partition (LMARP) problem, prove its NP-hardness, and develop polynomial-time approximation algorithms with guaranteed approximation ratios. Our experimental evaluation on the PEERING testbed demonstrates that ReOpt-optimized regional anycast achieves significant latency improvements, reducing 90th percentile client latency by 4.6-9.3% across diverse partitions compared to conventional regional anycast, demonstrating its effectiveness in adapting to real network conditions beyond static geographic constraints.

Deep Reinforcement Learning (DRL) promises adaptive control for future mobile networks but conventional agents remain reactive: they act on past and current measurements and cannot leverage short-term forecasts of exogenous Key Performance Indicators (KPIs) such as bandwidth. Augmenting agents with predictions can overcome this temporal myopia, yet uptake in networking is scarce because forecast-aware agents act as closed-boxes; operators cannot tell whether predictions guide decisions or justify the added complexity. We propose SIA, the first interpreter that exposes in real time how forecast-augmented DRL agents operate. SIA fuses Symbolic AI abstractions with per-KPI Knowledge Graphs to produce explanations, and includes a new Influence Score (IS) metric. SIA achieves sub-millisecond speed, over 200× faster than existing EXplainable Artificial Intelligence (XAI) methods. We evaluate SIA on three diverse networking use cases, uncovering hidden issues, including temporal misalignment in forecast integration and reward-design biases that trigger counter-productive policies. These insights enable targeted fixes: a redesigned agent achieves a 9% higher average bitrate in video streaming, and SIA’s online Action-Refinement module improves RAN-slicing reward by 25% without retraining. By making anticipatory DRL transparent and tunable, SIA lowers the barrier to proactive control in next-generation mobile networks.

In Federated Learning (FL) with over-the-air aggregation, the quality of the signal received at the server critically depends on the receive scaling factors. While a larger scaling factor can reduce the effective noise power and improve training performance, it also compromises the privacy of devices by reducing uncertainty. In this work, we aim to adaptively design the receive scaling factors across training rounds to balance the trade-off between training convergence and privacy in an FL system under dynamic channel conditions. We formulate a stochastic optimization problem that minimizes the overall Rényi differential privacy (RDP) leakage over the entire training process, subject to a long-term constraint that ensures convergence of the global loss function. Our problem depends on unknown future information, and we observe that standard Lyapunov optimization is not applicable. Thus, we develop a new online algorithm, termed AdaScale, based on a sequence of novel per-round problems that can be solved efficiently. We further derive upper bounds on the dynamic regret and constraint violation of AdaSacle, establishing that it achieves diminishing dynamic regret in terms of time-averaged RDP leakage while ensuring convergence of FL training to a stationary point. Numerical experiments on canonical classification tasks show that our approach effectively reduces RDP and DP leakages compared with state-of-the-art benchmarks without compromising learning performance.

The rapid expansion of modern wide-area networks (WANs) has made traffic engineering (TE) increasingly challenging, as traditional solvers struggle to keep pace. Although existing offline ML-driven approaches accelerate TE optimization with deep neural networks (DNNs), they often lack sufficient expressiveness and generalization on unseen traffic patterns or topologies, limiting their practicality. Inspired by the success of large language models (LMs), for the first time, this paper investigates their potential as general-purpose traffic planners. Our contributions are two-fold: (i) Theoretically, we show that pre-trained LMs can simulate the sequential decision processes underlying TE and, crucially, exhibit parallel reasoning capabilities, making them well-suited for the task; (ii) Practically, we present Lmte, a novel LM-driven TE framework that embraces these insights through efficient multimodal alignment and lightweight configuration generation, all while preserving the model’s original abilities. Extensive experiments demonstrate that Lmte matches top-tier performance on five datasets, achieving up to 15% better maximum link utilization (MLU) and consistently lower performance degradation across diverse scenarios, e.g., less than 5% with high traffic dynamics and link failures. Moreover, it achieves 10 to 100 times speedups over traditional TE solvers.