The Internet of Bodies (IoB) requires efficient transmission of correlated physiological signals under strict bandwidth constraints. We propose DiSC2, a bandwidth-efficient distributed semantic communication framework for cross-modal redundancy suppression, which combines distributed semantic encoders, cloud-side cross-modal attention, a hybrid task-anddistortion loss, and a lightweight quantization-aware edge encoder. Unlike independent waveform transmission, DiSC2 prioritizes complementary task-relevant information across modalities and shifts redundancy exploitation to the cloud. Experiments on PttPPG and ScientISST MOVE show that DiSC2 preserves downstream classification performance more effectively than SSC and DWT+LDPC baselines, especially under low-SNR and low-CBR conditions, while remaining suitable for resource-constrained wearable sensing.
Recent advances in programmable data planes (PDPs) have enabled the realization of in-network intelligence (INI), where machine learning models are directly executed within network switches. However, most existing INI systems assume static data distributions, overlooking concept drift, the temporal and contextual evolution of network traffic that leads to degraded inference accuracy and unreliable long-term performance. Although periodic retraining can mitigate drift, it introduces significant computational and communication overhead, especially without precise drift detection. To address this challenge, we propose LUCID, a lightweight unsupervised drift detection and selective update framework that operates entirely within programmable switches. LUCID monitors centroid deviations at decision tree (DT) leaf nodes to detect drift in real-time without relying on floating point or stateful computations, and it employs an anchor rule-based selective update to refresh only unstable rules rather than retraining the entire model. Tested on a software programmable switch, LUCID demonstrates effective drift detection with 77.0% improvement in accuracy and efficient model update 44.5% reduction in retraining overhead.
High-precision location in unstructured outdoor environments remains constrained by the physical complexities of the wild. Under a dense canopy, the Global Positioning System (GPS) fails to provide reliable coordinates, whereas existing radio-frequency sensing is often limited by the restrictive hardware dependencies of Channel State Information (CSI). We propose WiWildLoc, a system that uses the protocol-compatible Beamforming Feedback Matrix (BFM) to achieve device-free positioning using standard commercial hardware. To recover spatial details from compressed BFM data, we design a multimodal Siamese network that decouples spatio-temporal and angular features, yielding robust fingerprints that withstand severe multipath fading. Field evaluations in bamboo groves and highdensity forests demonstrate a mean positioning error of 0.191 m and an angular MAE of 7.84°. These results provide a scalable foundation for forestry inspection and field rescue, addressing the terminal precision gap in understory environments.
Modern short-video recommendation systems primarily rely on user-video interactions (e.g., sharing, rating, commenting) while neglecting nuanced emotional states, resulting in superficial engagement. To address this gap, we introduce MARS, an emotion-aware framework that dynamically aligns recommendations with viewers’ real-time affective states. We present two key contributions: First, leveraging multi-modal affective computing, MARS employs physiological signals from consumer-grade wearable biosensors (headphones, smartwatches, smartphones) to infer emotional context. Second, an emotion-aware graph convolutional network (EA-GCN) maps emotional trajectories continuously captured during viewing into a new edge embedding, enabling precise emotion-matched recommendations. This dual approach fosters adaptive personalization tailored to users’ evolving emotional needs. We validate MARS through a comprehensive study involving 70 participants and 30,000 user-video interactions. Results demonstrate $\mathbf{8 5 . 5 4 \%}$ classification accuracy in the circumplex settings, and $30.6 \%$ improvement in recommendation quality over state-of-the-art baselines. Our work advances affective computing in recommender systems by bridging psychological models with scalable machine-learning pipelines, offering a robust framework for emotionally intelligent personalization.
Discovering IPv6 router interfaces is crucial for network asset visibility and threat detection. However, existing technologies primarily focus on identifying address regions with a higher likelihood of yielding responses, generating / $\mathbf{6 4}$ prefixes within these regions as scanning targets while neglecting the utilization of existing active $/ 64$ prefixes. To address this gap, this paper proposes 6Hunter, an efficient IPv6 router interface discovery framework. 6Hunter performs pattern mining on known / 64 prefixes, generates / 64 prefixes as scanning targets based on the prefix patterns, and adopts a budget allocation method based on Bayesian smoothing mechanism for dynamic scanning. Experimental results demonstrate that with a budget of 100 million packets, 6Hunter outperforms Sweeper, Treestrace, FlashRoute, Yarrp6, and Scamper, achieving improvements of $\mathbf{1 2 0. 5 \% - 2 7 2 \%}$ in IPv6 address hit rate and $\mathbf{1 4 9 \% - 1 5 6. 1 \%}$ in IPv6 address discovery speed, respectively. Additionally, analysis of interface identifiers and ports of these IPv6 addresses indicates that a significant proportion of the discovered routers still face security vulnerabilities arising from privacy leakage and open ports.
Low-Power Wide Area Networks (LPWANs) support large-scale IoT connectivity but are capacity-limited, motivating concurrent transmissions to improve spectral efficiency. However, existing concurrent LPWAN solutions typically rely on a single base waveform or quasi-orthogonal waveforms, resulting in limited concurrency. In this paper, we propose Zadoff-Chu Random Access (ZCRA), a low-overhead framework for massive random access. ZCRA introduces a Zadoff-Chu-sequence-based waveform, which modulates multiple bits via cyclic shifts of Zadoff-Chu root sequences, exploiting their zero cyclic autocorrelation property. By allowing users to randomly select root sequences, ZCRA effectively mitigates inter-user interference without waveform pre-assignment based on low cross-correlation property. To resolve same-root collisions under low SINR, we further develop a constellation-set-match-based decoding scheme (CSMatch) that leverages spectral correlation structure to separate superimposed transmissions and remains robust to carrier frequency offsets and multipath channels. Extensive evaluation demonstrates that ZCRA-CSMatch significantly improves concurrency, achieving a lower symbol error rate compared with state-of-the-art collision resolution approaches, along with a $\mathbf{2. 1 3} \boldsymbol{\times}$ network throughput improvement over current concurrent systems and supporting 200 nodes (40 concurrency).
Modern mobile systems span smartphones, IoT devices, and edge nodes, evolving rapidly through frequent updates that expand their attack surface. These environments generate massive logs containing subtle security signals, such as mobile malware, rogue apps, insecure APIs, and compromised edge devices. Effective log anomaly detection (LAD) is therefore critical but challenging. Rule-based approaches and shallow models fail to generalize, while deep learning methods, though better at capturing log sequences, often suffer from high variance and fragile thresholding at scale, limiting their reliability for real-world mobile system security monitoring. To address the challenges, our core contribution is REMLog, a Robust Ensemble-based LAD framework designed to secure mobile systems. It unifies heterogeneous operational logs, from cloud control planes to sensor management platforms and mobile edge devices, into a single pipeline for scalable and reliable anomaly detection in complex mobile systems. To support high-volume log ingestion and lowlatency streaming analysis, REMLog integrates Apache Kafka as its ingestion backbone. Fundamentally, REMLog employs BAT, a bagging-based ensemble of EM-AT (Expectation Maximization-Anomaly Transformer) models with diversified parameterizations to significantly reduce decision variance and improve detection robustness. To address the difficulties of threshold sensitivity and manual calibration, we hereby implement EM-AT, an enhanced AT that integrates an EM procedure to automatically calibrate anomaly thresholds without manual tuning. In dynamic mobile environments, this allows the LAD model to adapt to varying noise profiles across diverse nodes. We evaluate REMLog on two widely used system log datasets, HDFS and OpenStack, which emulate realistic back-end infrastructures powering mobile and sensor service platforms. EM-AT achieves $F_{1}$-scores of $98.90 \%$ and 99.61 % on HDFS and OpenStack, respectively, while REMLog further improves performance to $99.99 \%$ on both datasets. These results show that REMLog delivers accurate, scalable, and robust LAD suitable for proactive security monitoring in mobile systems.
In LoRaWAN, users are allowed to customize the center frequency of channel. This mechanism provides more flexibility but also introduces non-negligible interference for packet reception. This paper focuses on Overlapping Channel Interference (OCI), which arises when multiple users simultaneously transmit on partially overlapping channels. Existing methods proposed for interference in a single channel concurrently demodulate collided packets by inferring the chirp’s onset and offset times. However, the collided chirps in OCI exhibit different center frequencies, leading to unpredictable chirp’s onset and offset times which makes OCI particularly challenging to resolve. To resolve it, we propose Hole, a novel method to reliably receive target packet by exploiting the difference in chirp’s onset and offset times instead of inferring them. Hole adopts signal elimination to construct identifiable energy losses for different chirps. Since the onset and offset times of target chirp are stable, its energy loss is deterministic while that of the interfering chirp is random. By comparing energy loss, we identify target chirp under OCI. However, in practice, the channel noise reduces the stability of energy loss. Therefore, we first theoretically analyze the optimal position of signal elimination and then propose noiseaware elimination setting. Furthermore, we also propose target packet recognition method by leveraging difference in packet’s center frequency offset. The experiments in real LoRa network present that Hole improves the Packet Reception Rate (PRR) by up to $7.5 \times$ compared with existing methods.
Real-time anomaly detection in Open Radio Access Networks (O-RAN) demands high accuracy, low false alarms, and computational efficiency for resource-constrained edge deployment. Traditional methods struggle with computational overhead, inconsistent cross-domain performance, and suboptimal feature representations that miss subtle attacks on O-RAN's open interfaces. We present CoMeT-Net (Consensus Memory Template Network), a framework achieving state-of-the-art detection through three innovations: (1) structured memory banks enabling template-based consensus voting with O(N · C) complexity; (2) adaptive gating that downweights ambiguous features as a learned noise filter; (3) contrastive alignment unifying feature learning and classification. Deployed in O-RAN infrastructure via edge servers and Near-RT RIC xApp, CoMeT-Net enables dynamic threat mitigation through PRB throttling and RRC connection release. On network traffic datasets, CoMeT-Net achieves 99.35
Unmanned Surface Vehicles (USVs) play a crucial role in maritime surveillance, resource exploration, and disaster response. However, their communication range and computing capabilities are limited, which causes data transmission delay and an inability to execute complex computing tasks. Existing studies focus on local computing or small-scale USV collaborative computing, but they suffer from insufficient resource pooling and ineffective computing capabilities for dependent tasks. To address these issues, this paper proposes an air-sea collaborative computing method for open-sea USVs. Specifically, we first propose a USV clustering mechanism to group USVs dynamically to better organize their computing resources. Secondly, we construct an Unmanned Aerial Vehicle (UAV)-assisted multi-hop link model to enable stable data transmission among USV clusters. Based on this, we construct an air-sea collaborative computing model that uses latency and energy as objectives and considers task dependency, maximum tolerable delay, resource capacity limits of USV clusters, and UAV relays as constraints. Finally, we improve the multi-agent deep deterministic policy gradient algorithm by integrating cluster-aware state representation and dynamic resource allocation mechanisms, and use it to solve the proposed model. Simulation results show that the proposed model outperforms baseline methods significantly in terms of latency, energy, and task completion rate.
Coupling conflicts in sensor-cloud environments become particularly severe when concurrent user demands compete for a limited pool of physical resources. Most existing cloudand edge-oriented scheduling approaches are built on assumptions of request independence and one-to-one allocation, which limits their ability to handle multi-user coupling effects arising from shared-resource competition. In this work, we formalize the scheduling problem and establish its NP-hardness via a polynomial-time reduction from the Partition problem. To tackle this challenge, we develop a low-coupling edge scheduling scheme that integrates dual-buffer queue management with an improved Kuhn-Munkres (KM) matching strategy. The method adaptively consolidates redundant service requests and computes nearoptimal request-resource mappings with a time complexity of $\mathbf{O}\left(n^{3}\right)$ and an approximation ratio of $\mathbf{O}(\ln m)$, where m denotes the number of requests. Extensive experiments on a real testbed further demonstrate consistent reductions in scheduling delay and improvements in resource utilization when compared with representative baseline methods. These results demonstrate the proposed approach’s capability to effectively alleviate couplinginduced performance degradation in sensor-cloud systems.
Collisions among LoRa packets are common in dense low-power wide-area networks and remain a key bottleneck for reliability and capacity. While existing multi-packet reception schemes can decode overlapped LoRa frames at gateways, they typically rely on preambles, known PHY parameters, or reception diversity, and rarely provide blind waveform-level separation from raw I/Q samples. This paper investigates blind separation of two simultaneously active LoRa sources at a dual-antenna receiver, which can be instantiated as a multi-antenna gateway, a testbed node, or a USRP-based spectrum monitor. Using a USRP B210, we collect 2400 real indoor/outdoor two-source collision samples under four source-array distances $(5 \mathrm{~m}, 10 \mathrm{~m}, 15 \mathrm{~m}$ and 20 m) and three angle-of-arrival separations ($30^{\circ}, 45^{\circ}$ and 60°). We formulate a dual-channel complex STFT-domain mixing model and propose a lightweight multi-channel ResNet masking network that exploits both CSS time-frequency structure and inter-antenna phase cues, trained end-to-end with permutation invariant training (PIT). On 240 held-out mixtures, the proposed separator achieves 17.4 dB mean SI-SDR and 1.72% mean pseudo-SER, comparable to a U-Net baseline and more robust than Conv-TasNet at very low SNR. Across input SNR from -10 dB to 15 dB, the model maintains pseudo-SER around 2% once SNR $\geq 0 \mathrm{~dB}$, suggesting its potential as a practical physicallayer collision recovery module for LoRa gateways, testbeds, and spectrum monitors.
Inferring routing paths in overlay networks is challenging because overlay encapsulation hides physical paths from tools like traceroute, while end-to-end metrics such as round-trip time (RTT) alone are insufficient for reliable inference. We propose leveraging clock offsets alongside RTTs as a two-dimensional latency space, and formulate a latency consistency condition over node triplets to infer routing paths in overlay networks. A noiseadaptive threshold is introduced to tolerate measurement noise and ensure robust inference. Evaluations via simulations, testbed emulations, and real-world cloud deployments demonstrate the effectiveness of our approach.
Quantum entanglement routing in Low Earth Orbit (LEO) satellite networks faces critical challenges from dynamic topology, decoherence and limited quantum resources. We propose Quantum entanglement Routing with Integrated Selection and Enhancement strategy (Q-RISE), an optimisation algorithm that jointly addresses fidelity thresholds, communication delays and resource consumption under coherence time constraints. QRISE integrates a Markov-enhanced Cascaded Swap Fidelity Model (MCSFM) for accurate fidelity estimation and a Minimum Purification Rounds (MPR) mechanism for resource-efficient entanglement enhancement. Simulation results demonstrate that QRISE outperforms Q-CAST, the strongest baseline, by reducing average waiting time by up to $46 \%$, decreasing quantum resource utilisation by up to $67 \%$ and maintaining high end-to-end (E2E) fidelity in networks of up to 90 nodes. These results confirm QRISE as a scalable and reliable entanglement routing solution for future quantum LEO satellite constellations.
The synergy between 5G and Virtual Reality is transforming the entertainment landscape, with Open Radio Access Network (O-RAN) offering a flexible architecture for private 5G deployments. This paper introduces a comprehensive testbed to analyze Multi-User VR (MUVR) gaming performance over O-RAN. We identify three key factors that severely degrade the end-user Quality of Experience (QoE): prolonged tail latency, uneven bandwidth allocation, and unfair video quality. To address these problems, we present Octopus, a real-time interactive streaming system designed for MUVR over O-RAN. Octopus implements a bidirectional coordination mechanism that closes the loop between Adaptive Bitrate (ABR) streaming and RAN scheduling. By leveraging the O-RAN RAN Intelligent Controller (RIC) framework, the system achieves cross-layer optimization while maintaining architectural flexibility. Experimental results demonstrate that Octopus significantly outperforms state-of-the-art VR streaming solutions, reducing the 99.99th percentile latency by $48.45 \%$, increasing average bitrate by $23.60 \%$, and improving streaming smoothness and max-min fairness by $63.16 \%$ and $104.34 \%$, respectively.
LoRa is a popular low-power wireless technology in the Internet of Things (IoT), but direct interaction with consumer devices is hindered by protocol incompatibility. Existing crosstechnology communication (CTC) solutions for Wi-Fi receivers often require additional gateways or transmitters. We propose a LoRa-to-Wi-Fi CTC technique that enables smartphones to directly receive and decode LoRa transmissions using only their native hardware capabilities. Our key innovation is utilizing the spectrum scanning functionality in commercial mobile Wi-Fi chipsets for LoRa packet decoding. A signal processing pipeline combining frequency-domain analysis and lightweight neural-network-based symbol classification reliably recovers LoRa packets under commodity hardware constraints. Our prototype, implemented on a Google Pixel 5, demonstrates robust one-to-one LoRa-to-Wi-Fi communication with commercial LoRa tags, enabling smartphones to act as ad hoc LoRa listeners and simplifying infrastructure-free IoT data collection.
Pervasive encryption has fundamentally reshaped modern communication networks, reducing the effectiveness of traditional payload-based traffic analysis and elevating encrypted traffic classification into a core network sensing challenge. Accurate identification of encrypted traffic types is important for network monitoring, security enforcement, and resource management, yet it remains difficult due to the loss of semantic visibility and the diversity of network behaviors across applications and devices. In this paper, we introduce Hilo-MAE, an efficient pre-training framework that captures hierarchical local-global dependencies in encrypted flow data. From a networking perspective, encrypted traffic exhibits multi-scale patterns within the flow prefix, where fine-grained local byte/packet patterns and coarse-grained flow-level context jointly determine traffic semantics. HiLo-MAE leverages a Hierarchical Local-Global (HiLo) attention mechanism within a masked autoencoder architecture to model these complementary patterns, enabling robust flow representations under encryption. We evaluate HiLo-MAE on seven encrypted traffic datasets spanning enterprise, mobile, and IoT network scenarios. Experimental results show that HiLo-MAE consistently outperforms the evaluated baselines, achieving improvements of up to $26.5 \%$ in classification accuracy over the strongest compared method while maintaining favorable computational efficiency. Results across multiple benchmark datasets and few-shot settings further suggest that the learned representation can reduce reliance on large labeled datasets, making HiLo-MAE a promising candidate for scalable deployment in real-world network monitoring systems.
The availability of low-cost, low-power UltraWideband (UWB) ranging radios in the past decade has enabled applications that can benefit from accurate distance time-series measurements. These applications benefit from UWB accurate wireless ranging capability, which is robust in both indoor and outdoor environments. However, adopting UWB technology in constrained wearable systems requires confronting the challenge of high radio power consumption. We present the design of a wrist-worn wearable system that combines UWB two-way ranging with Bluetooth Low Energy and an ultra-low power accelerometer to achieve high ranging accuracy while consuming significantly less power than a UWB-only ranging solution. In this work, we focus on applications that require measuring human-to-human or human-to-object distance time-series over long time intervals. For such continuous ranging applications with specific energy efficiency requirements, we demonstrate that the average UWB power consumption can be minimized by using an adaptive UWB ranging policy learned through reinforcement learning. We evaluate our design with a new hardware prototype, RangeTag, which is based on the Qorvo DWM1001 module. Our design enables energy cost and ranging accuracy trade-offs for various applications. Our evaluation shows that, for a range of ultra-low power scenarios, ranging accuracy can be improved by 20% to 40% on average compared to UWB alone.
The deployment of deep neural networks (DNNs) on resource-constrained edge devices demands a rigorous tradeoff among inference accuracy, latency, and energy efficiency. While recent architectural innovations like Global Filter Networks (GFNet) offer efficient frequency-domain token mixing, deploying them on embedded platforms remains challenging due to static computational graphs that fail to adapt to varying input complexity. To address this, we propose EAE, a holistic Efficient image Analysis framework for Edge devices that synergizes spectral-domain model compression with a hardwareaware dynamic inference mechanism. First, we introduce a novel frequency-domain knowledge distillation technique that regularizes student feature norms and orientations against a teacher network, achieving significant parameter reduction while preserving global receptive fields. Second, we dismantle the “exit-at-every-layer” paradigm common in dynamic networks, which we demonstrate creates prohibitive synchronization overheads for lightweight backbones. Instead, EAE employs a strategic, interval-based exit placement optimized for the specific latency characteristics of embedded GPUs. Furthermore, we propose a heterogeneous orchestration strategy that offloads lightweight gating and uncertainty estimation to the CPU, thereby minimizing GPU kernel interruptions. Extensive experiments on NVIDIA Jetson Orin Nano and AGX Orin platforms across three standard benchmarks demonstrate that EAE achieves an average inference speedup of $1.62 \times$ and $19.9 \%$ energy savings compared to state-of-the-art dynamic networks, with gains extending to $1.96 \times$ and $59.3 \%$ against the vanilla GFNet, all while maintaining competitive accuracy. Our code and dataset will be released at: https://github.com/GG-Bond-Ultra/EAE.
