Modern data center networks underpin latency-critical services such as cloud computing, AI workloads and emerging 6G applications, yet even brief switch or link failures can cascade into severe service disruptions. Existing fault detection systems face limitations: control plane centric methods incur high latency (seconds to minutes) and suffer control channel overload, while data-plane-centric approaches often rely on intrusive probing, limiting scalability. We present MPulse, a multi-dimensional fault detection and recovery framework. Our primary idea is to differentiate the roles of monitoring switches and transmission switches, shifting fault management to the data plane through two key innovations: collaborative multidimensional monitoring and a decentralized voting mechanism. Following this, we enable switches to collaboratively collect real-time status, historical data, and topological state through periodic information exchange, allowing each node to aggregate global multi-dimensional insights. Then, a distributed voting process is designed to allow switches to independently determine fault types (link/switch failure) and locations based on integrated data dimensions, eliminating dependencies on centralized control. Implemented on Intel Tofino switches with minimal overhead, MPulse achieves sub-millisecond fault localization (<1 ms), 324.8× faster detection than existing solutions, and near-100% accuracy under congestion. Experimental results validate its effectiveness and low operational costs.
Remote Direct Memory Access (RDMA) enables high-performance communication through kernel bypass and zero-copy transfers. However, as data center clusters scale up, the number of connection grows explosively, quickly exhausting the limited on-chip memory of RNICs and triggering frequent cache misses and off-chip memory accesses, which severely constrains connection scalability and becomes a major deployment bottleneck. Existing connection-sharing schemes reduce connections counts but largely ignore head-of-line (HOL) blocking and performance interference caused by multi-flow contention on shared connections. In this paper, we present Cadence, a lightweight, fine-grained scheduling–driven RC QP sharing mechanism implemented as a user-space middleware. Cadence integrates three key components: (1) a WQE splitter that segments large messages to mitigate HOL blocking; (2) a WQE scheduler that identifies and differentially schedules heterogeneous traffic; and (3) a WQE monitor that infers RNIC cache pressure and adaptively tunes the scheduling window via closed-loop feedback. NS-3 evaluation results show that Cadence significantly reduces flow completion time (FCT) while sustaining high peak throughput. Compared to RC, Cadence reduces average FCT, P99 FCT, and average short-flow FCT by approximately 34%, 21%, and 56%, respectively, while improving throughput by 7.8%.
Network operators utilize in-network bandwidth allocation strategies to ensure fairness among competing flows. Existing approaches mainly rely on two mechanisms deployed in switches: packet scheduling algorithms and probabilistic packet dropping. However, their effectiveness is constrained by the limited number of physical priority queues and the hardware overhead of arithmetic operations. The former leads to high packet loss, resulting in low link utilization, while the latter impairs bandwidth allocation accuracy.In this paper, we propose AGILE, a novel bandwidth allocation strategy that enhances network fairness while maintaining high link utilization. Despite the limited flow visibility in switches, AGILE performs real-time global flow state tracking and optimal per-flow fair share computation to achieve max-min fairness, dynamically allocating residual bandwidth to maximize link utilization. We implement a prototype of AGILE on Intel Tofino switches. Our experimental results show that AGILE attains an average Jain’s Fairness Index (JFI) of 0.996. Compared to state-of-the-art approaches, including Cebinae and AHAB, AGILE reduces relative bandwidth allocation errors by 45.7%–86.3%, improves link utilization by 5%, and decreases link utilization fluctuations by 37%–75%.
Retrieval Augmented Generation (RAG) greatly mitigates hallucinations in large language models by grounding responses in external documents, yet its adoption on mobile devices is hindered by two intertwined concerns: (i) security, because poisoned passages cached by adversaries can drive the model to generate harmful outputs, and (ii) efficiency, because deeper retrieval and wider uplinks consume limited battery and spectrum. Existing RAG security studies overlook bandwidth and power budgets, whereas mobile edge offloading literature assumes benign knowledge bases. Bridging this gap, in this paper we introduce the Security Cost Ratio (SCR) as a unified metric that rewards lower attack-success rate per joule expended, and we design an optimization problem to jointly control retrieval depth, bandwidth and transmit power in the mobile edge computing environment. To tackle the non-convex optimization problem, we derive closed-form KKT conditions and an epigraph transformation that attain the global optimum for retrieval depth. We then utilize a quadratic-transform fractional programming technique in an alternating framework that converges for bandwidth and transmit power allocation. Simulations across various system parameters and environmental settings demonstrate that the proposed algorithm achieves up to 1.8× higher SCR and 50% lower attack success rate, while maintaining rapid convergence.
With the rapid advancement of communication technologies, wireless data transmissions have become pervasive. However, the open nature of the wireless medium makes these transmissions vulnerable to attacks, such as malicious detection and eavesdropping. Importantly, the prerequisite for these threats is the attackers’ ability to detect the existence of legitimate transmissions. Covert communication (CovCom), a physical-layer security (PLS) technique designed to conceal legitimate transmissions from detection by attackers, serves as a fundamental security measure unlike other PLS schemes. Nevertheless, most existing CovCom schemes rely on covering signals that are independent of the covert signals themselves. This independence allows attackers to enhance their signal processing capabilities to separate the covert signal from the covering signal, ultimately extracting the original signal. Furthermore, the necessity for dedicated covering signal generation incurs additional power and hardware costs. To address these deficiencies, we propose a novel scheme called Generative Covert Communication (GCC), aimed at achieving covert data transmissions. GCC leverages interactions among covering signals; by adjusting their power and/or phases, a desired secret signal intended for the covert receiver (Rx) can be produced. Since the covert signals are derived from covering signals, they become coupled, and hence prevent attackers from decoupling them and detecting the existence of, as well as the content carried by, the secret transmissions. We design two GCC schemes: Power Control GCC (PC-GCC) and Phase Control GCC (${\mathbb{F}}$C-GCC). We utilize (1) the Universal Software Radio Peripheral (USRP) platform for experimental validation of GCC’s feasibility and (2) comprehensive MATLAB simulations to demonstrate its effectiveness in achieving covert and secure data transmissions.
This demo showcases the use of distributed acoustic sensing (DAS) for urban traffic monitoring. We remotely connect to a field-deployed DAS interrogator to stream live vibration traces from a fibre-optic cable; however, due to time-zone differences and potential connectivity limitations, the demonstration will be primarily supported by synchronized video recordings. Using pre-trained models, the demo will illustrate real-time detection and classification of pedestrians, cyclists, and vehicles, as well as estimation of their direction of movement and speed from DAS signals. By presenting a practical DAS deployment under realistic traffic conditions, this demo highlights the potential of DAS as a scalable tool for urban and active travel monitoring.
Collaborative vision-based perception is a pivotal technology for intelligent transportation systems, enabling vehicles to transcend individual sensing limits and enhance situational awareness through information exchange. Despite its significance, existing literature lacks a rigorous analytical framework to evaluate such systems while accounting for the inherent spatiotemporal correlation of vehicular networks. In this paper, we establish a novel stochastic geometry-based model that characterizes vehicular distributions as a Poisson line Cox process (PLCP). To address performance degradation due to sensing-to-transmission latency, we propose a spatio-temporal modeling framework that integrates direct sensing with predictive enhancement. Furthermore, we introduce the mean effective area fraction as a performance metric, which incorporates a temporal validity function to capture the decay of information freshness. Within this framework, we derive analytical expressions for the joint rate coverage probability by quantifying the cross-slot interference correlation induced by vehicle mobility. The accuracy of our theoretical derivations is validated via Monte Carlo simulations. Our results provide critical system-level insights into the interplay between vehicular density, communication reliability, and perception efficacy. This study represents a first step in integrating spatio-temporal correlation and information validity into the evaluation of collaborative vehicular perception.
Typically, Reinforcement Learning (RL) agents require complete knowledge of the state of the system to determine the optimal action. However, when an agent cannot directly observe its environment and must rely on data from a remote node, communication impairments and delays can degrade the learning process and the resulting control policy. Recently, numerous frameworks have been proposed to handle RL with imperfect feedback, but they require significant communication overhead. To overcome these limits, we propose a novel architecture, named Homomorphic Robust Remote Reinforcement Learning (HR3L), that enables the training of remote RL agents collecting observations through a non-ideal wireless channel. The proposed solution entails a transmitter that observes the system state and encodes it in meaningful messages, following a goal-oriented approach, and a receiver that decodes these messages and performs actions to maximize its long-term discounted reward. Crucially, HR3L does not require the exchange of gradient information across the channel, accelerating the training process and reducing the communication overhead with respect to state-of-the-art solutions. Experimental results show that HR3L significantly outperforms baseline methods in terms of sample efficiency and adapts to packet losses and delayed transmissions without retraining.
In wired networks, Medium Access Control (MAC) spoofing is a well-known and extensively studied class of network security attacks in which an adversary impersonates a legitimate host by falsifying its MAC address. Despite existing defense standards, such attacks remain relevant in contemporary networks. MAC spoofing can enable Man-In-The-Middle (MITM) attacks, allowing an attacker to intercept and inject traffic while posing as the victim whose MAC address has been compromised. Despite its conceptual simplicity, implementing a fully transparent MAC spoofing MITM attack is non-trivial. Existing techniques fail to maintain transparency from the victim’s perspective and may disrupt connectivity. While traffic interception is relatively straightforward, injecting traffic on behalf of the victim without affecting the victim’s normal operation is more challenging. Achieving such transparency typically requires custom code and specialized libraries that may lack portability across platforms or operate in user space rather than kernel space, thereby imposing performance limitations. This demo presents an implementation of a MAC spoofing-based MITM attack that enables transparent interception and injection of packets while impersonating the victim, without requiring custom code. The approach operates entirely in kernel space, leveraging Linux networking names-paces, and ensures complete transparency to both the victim host and its communication endpoint.
We study the problem of constraint-aware combinatorial multi-armed bandits (CMAB), a novel extension of combinatorial semi-bandits that maximizes total rewards while adhering to long-term cost constraints. In each round, the environment generates stochastic rewards and costs for each arm from unknown distributions. The learning agent selects a combinatorial set of arms that collectively gain rewards and incur costs, which are (partially) observed as feedback to improve future decisions. While prior work has focused on minimizing regret, it often overlooks cost constraints and diverse reward structures across applications. To address this, we introduce a general framework, constraint-aware CMAB with versatile reward functions (C2MAB-V). Unlike existing works that solve a difficult discrete constrained optimization (CO) problem to control constraint violation, we propose bypassing the original discrete CO problem and using a novel relaxation and rounding (RR) approach, which solves a relaxed continuous CO problem with improved approximation guarantees and computational efficiency. A new challenge arises since RR is originally designed for offline CO problems, and due to the flexibility of choosing the relaxation function F and the rounding procedure σ, it is unclear how to guarantee that RR can achieve low regret and low violation simultaneously. In this paper, we are the first to discover unified RR conditions and, equipped with these conditions, we prove that C2MAB-V has the following nice properties: sub-linear regret, fast diminishing violation, and good computational efficiency. We demonstrate the generality of our RR conditions by applying C2MAB-V to diverse applications with concrete (F,σ), i.e., mobile crowdsensing, network routing, and multi-LLM selection, with varying reward functions (e.g., linear, conjunctive, disjunctive, submodular) and feedback models (e.g., semi-bandit, cascading). Extensive experiments on synthetic and real-world datasets, including evaluations with nine LLMs on practical tasks, confirm that C2MAB-V outperforms baselines across these applications.
Large Language Models (LLMs) that employ extended Chain-of-Thought (CoT) decoding incur substantial inference-time computation, raising concerns about service availability under adversarial workloads. Prior work such as ExtendAttack demonstrates that carefully crafted prompts can substantially prolong model inference; however, it overlooks the practical efficiency of such attacks from the attacker’s perspective. In this paper, we analyze resource depletion attacks through the lens of attack efficiency and identify a key limitation of static prompt obfuscation: excessive input inflation increases the attacker’s upload cost without corresponding gains in server-side computation. We propose NA-ExtendAttack, a network-aware attack framework that adapts prompt obfuscation to the attacker’s bandwidth conditions to improve the amplification ratio, defined as generated tokens per uploaded byte. Experiments on a vLLM-based non-streaming inference service with Qwen3-8B demonstrate that, under bandwidth-constrained settings, NA-ExtendAttack achieves up to 54% higher amplification ratio compared to ExtendAttack. Our analysis provides insights into the network-compute interactions that determine attack feasibility and offers guidance for designing efficiency-aware defenses for GenAI inference services.
The creation and deployment of honeypots traditionally require significant manual effort and specialized expertise, limiting their adaptability and diversity. Recent advances in Large Language Models (LLMs) enable new forms of automation through natural language interfaces. This work introduces Chat2Pot, a system that leverages LLMs to dynamically generate diverse, Docker-based honeypots from natural-language descriptions. By automating the generation of technical artifacts, Chat2Pot significantly reduces manual effort while enabling scenario-specific customization. The system integrates validation, orchestration, and containerized deployment to ensure robustness and safe automation. This live demonstration showcases the on-demand generation and deployment of honeypots based on user-provided descriptions of target environments.
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.
The proliferation of geo-distributed AI requirements necessitates extending Remote Direct Memory Access (RDMA) to Metropolitan Area Networks (MANs). However, a structural mismatch impedes this expansion: packet reordering and extended Round-Trip Times (RTTs) render RDMA mechanisms ineffective, while inherent traffic micro-bursts create severe contention for scarce wide-area bandwidth. To bridge this gap, we introduce Lossless New Metropolitan Area Network (LN-MAN), a controller-orchestrated architecture that adapts MAN infrastructure to satisfy strict RDMA requirements. LNMAN ensures packet ordering through SRv6-based deterministic forwarding and achieves RTT-independent congestion avoidance via bandwidth reservation based on network slicing. Furthermore, we repurpose wide-area flow control as a fine-grained lossless traffic shaping mechanism for lossless micro-bursts suppression that reduces bandwidth headroom. The controller orchestrates the entire service lifecycle, enabling production-grade resilience through dynamic optimization and automated failover. Evaluations on China Telecom New MAN testbed demonstrate that LNMAN supports collaborative training with less than 3% efficiency degradation while reducing bandwidth requirements by over 10×. This reduction in network resource requirements significantly lowers deployment costs and increases service capacity. Consequently, the deployment of LNMAN enables network operators to deliver stable wide-area RDMA services, with deploying costs remaining well within acceptable limits.
The integration of non-terrestrial networks into future 6G systems is a key enabler for global coverage, but it introduces significant challenges due to heterogeneous propagation delays, dynamic topologies, and limited onboard processing capabilities. This paper analyzes Central Unit (CU)-based multi-connectivity architectures for Open RAN space–terrestrial networks, with a focus on latency-aware operation and Quality of Service (QoS) support. We introduce and compare Single-PDCP and Multi-PDCP designs, characterized by the number of PDCP instances managing the data transmission, and evaluate their impact on end-to-end latency and compliance with 3GPP QoS Identifier (5QI) latency requirements. The feasibility of distributing RAN functions across space and terrestrial segments is assessed, highlighting the associated functional split and latency tradeoffs. Finally, we establish a mapping between PDCP architectures and multi-connectivity strategies and identify open challenges for efficient O-RAN–based space–terrestrial integration.
Collaborative perception enhances situational awareness among vehicles by enabling information sharing across agents. However, network-induced delays reduce temporal consistency and introduce spatial divergence in shared observations, posing a critical challenge in distributed, delay-sensitive vehicular networks. Although delay alignment mitigates part of the effect, alignment errors and inherent multi-view discrepancies often lead to inconsistencies when synchronizing delayed data to a common timestamp. Crucially, these inconsistencies can accumulate during fusion, propagating cascading errors into downstream trajectory prediction. To address this, we propose STCC, a prediction-oriented collaborative perception framework based on Spatio-Temporal Co-Sensing Calibration. STCC adopts a temporal-spatial cascade design: a history-aware temporal alignment module captures dynamic alignment patterns to compensate for variable delays, while an Interaction-Aware Uncertainty Quantification scheme explicitly models environmental interaction risks to correct cross-view representation divergence. This joint calibration ensures consistent multi-agent fusion and reliable downstream prediction under network-induced delay. We evaluate STCC on large-scale public cooperative perception benchmarks. Experimental results show that STCC consistently outperforms state-of-the-art methods across diverse delay conditions, significantly reducing the Average Displacement Error (ADE) and Final Displacement Error (FDE), demonstrating superior robustness and consistency.
The proliferation of heterogeneous services in Society 5.0 demands highly adaptive network slicing, yet the dynamic and unlabeled nature of traffic makes predefined service-to-slice mapping inefficient. This paper proposes a co-evolutionary learning framework to address autonomous slice matching without prior labels. By integrating online unsupervised clustering for real-time service inference with dual-timescale deep reinforcement learning for coupled decision-making, our method jointly optimizes slice-level resource orchestration and flow-level scheduling. Simulation results demonstrate that this approach significantly saves long-term system cost and improves task success rate under non-stationary traffic conditions.
This paper proposes a model-free unsupervised learning approach in pinching antenna systems (PASS) that directly learns antenna positions and power allocation from site-specific environmental information without explicit channel estimation. A Policy deep neural network (DNN) is designed to learn the antenna positions and power allocation. To train the DNN to maximize the spectral efficiency (SE) under the condition of unknown channel coefficients, a Reward DNN is trained to learn the data rate fed back from users, which can guide gradient back-propagation for training the Policy DNN. By designing the two DNNs as graph neural networks that can leverage the permutation equivariance (PE) property, the proposed method achieves efficient learning. Simulation results show that by introducing the PE property and proactively suppressing interference with blockages, the proposed GNN can achieve higher SE than time-division multiple access with reduced feedback overhead than fully-connected neural networks.
