Integrated Sensing and Communication (ISAC) enables ubiquitous Human Activity Recognition (HAR) but raises severe resource and raw data-localization challenges by centralizing raw sensing data. To address this, we propose an Explainable Split Federated Learning (SplitFL) framework for effective ISAC beamforming. Our architecture partitions the neural network at the edge, compressing raw signals into low-dimensional, noninvertible “smashed data” before transmission. Experimental results on the 6G Internet of Things (IoT) Intelligent Management Dataset demonstrate a $5.0 \times$ reduction in uplink payload and 99.95 % computation offloading to the server, resulting in an ultra-lightweight client footprint (0.36 KB). The model achieves 97.0 % accuracy in small-cluster settings, while our novel Split Grad-CAM mechanism validates that decisions are driven by physical channel attributes rather than artifacts. Our method, SplitFL, is an efficient and transparent paradigm for nextgeneration HAR systems empowered by ISAC and the state-of-the-art distributed learning technology.

To address data scarcity and computational constraints in medical imaging, we investigate a deep learning (DL)-based framework for automating nail disease diagnosis. A convolutional neural network (CNN) is combined with explainable artificial intelligence (XAI) techniques to enable both accurate classification and interpretable decision-making for nail diseases associated with nutritional deficiencies. Multiple CNN architectures are evaluated, achieving classification accuracies exceeding 93%. The integration of CNN and XAI provides visual explanations that enhance model transparency while maintaining strong diagnostic performance under limited training data and resource-constrained conditions. In addition, this work highlights the limitations of current XAI methods and discusses several research directions focused on reliability, vulnerability, and robustness analysis to support the trustworthy deployment of AI-assisted diagnostic systems.

This paper presents a preliminary study on LLM workflows with agent skills for the identification and classification of security vulnerabilities. We compare a plain LLM baseline with a lightweight staged workflow composed of three reusable skills: Context Skill, Reasoning Skill, and CWE Mapping Skill. Given a Potential vulnerability, the workflow first organizes the relevant code context, then performs structured security reasoning, and finally predicts both a vulnerability label and a CWE category. We evaluate the proposed workflow on a small benchmark of code examples and compare it with a single-prompt baseline using the same underlying model. The results show that the skill-based workflow yields more consistent identification and classification decisions and produces more interpretable rationales. Our findings suggest that even a simple, modular design can make LLMs more effective for software security tasks. This paper offers a practical step toward incorporating agent-style methods without the complexity of full multi-agent systems.

The burgeoning interest in Large Language Models (LLMs) has demonstrated significant utility across diverse domains, including Intrusion Detection Systems (IDS). However, they are prone to severe security threats, including Prompt Injection Attacks (PIA), where adversaries subtly manipulate input prompts or input logs to mislead the LLM into producing erroneous or inappropriate output, undermining its prompt quality. The current literature fails to account for PIA in real-time scenarios where LLMs are applied to IDSs. Given their utmost demand for security, optimal functionality, and token exhaustion, PIA undermines the overall practicability of the security application. To tackle this issue, we first evaluate the efficacy of LLMs for IDS, then examine the impact of a novel recursive PIA that poisons the LLM with unauthorized, repetitive tasks, thereby undermining real-time IDS and achieving a stealthy, wasteful, exhaustive, and time-evasive attack success rate. To defend against this recursive PIA, a Resource-Aware Prompt Injection Defense (RAPID) is proposed in this paper, with the core aim of sanitizing training data (network traffic) input being fed into the LLM, enforcing optimal token utility, and real-time constraints of the LLM-enabled IDS. RAPID distinguishes itself from extant PIA defenses by uniquely targeting recursive adversarial patterns with a lightweight, pre-execution sanitization layer that simultaneously restores semantic integrity and enforces strict computational bounds, thereby neutralizing the latency-driven evasion windows and resource-exhaustion vulnerabilities inherent in deploying LLMs for real-time IDS. Our empirical evaluation results not only verify the viability of RAPID but also demonstrate its ability to secure LLM-enabled IDSs against PIA.

The Internet of Medical Things (IoMT) is revolutionizing modern healthcare by enabling seamless connectivity among medical devices, which supports real-time patient data collection and analysis. While these advancements improve clinical outcomes, IoMT devices are increasingly exposed to cyber threats because of their high-value data and resource-constrained nature. In this paper, we address these challenges by comparing and evaluating different lightweight Deep Learning (DL) and Machine Learning (ML) techniques enabled intrusion detection systems. Our results demonstrate that the proposed approach can achieve precise, real-time, and cost-effective detection of multiple threats, offering a significant step forward in securing IoMT infrastructures.

The study proposes a two-stage decentralized energy management system for multi-microgrids, includes peer-to-peer trading and resilient network reconfiguration. Physical constraints such as radiality, voltage stability, and line capacity are implemented in first stage to ensure safe operation in both normal and faulty conditions. In second stage, market-clearing mechanism enables supply-demand bidding and zonal cost allocation for multi-bilateral trades while maintains grid and marginal pricing. Renewable uncertainty is modeled using an upper-quantile approach from historical data to balance robustness and economic efficiency without probability distributions. Financial incentives are used in incentive-based demand response following faults to shift or curtailment loads. The model is implemented on a modified IEEE-33 bus system with three microgrids and 36 residential households equipped with photovoltaic panels, wind turbines, battery energy storage, and electric vehicles. Simulation results show that under normal conditions, grid reliance for residential loads is reduced by 43.66%. During grid and line outages, demand falls by 16.33% while grid usage increases modestly to 53.16%. When distributed generators fail, peer to peer energy sharing within the faulty zone rises by a factor of 2.5, supported by battery and electrical vehicle discharges. The proposed framework thus enhances resilience, lowers operating costs, and strengthens local energy self-sufficiency through coordinated P2P trading, flexible storage, and fault-tolerant scheduling.

Edge Artificial Intelligence (i.e., Edge AI) has emerged as one key technology for next-generation smart applications in Internet of Things (IoT) and Cyber-Physical Systems (CPS), enabling low-latency, real-time decision making near the data source. To support workforce development, leading tech companies such as Texas Instruments, NVIDIA, and Intel have released their career-oriented educational materials focused on edge AI. However, college-level education continues to lag in offering up-to-date, practice-oriented curricula that reflect industry needs. Furthermore, existing academic courses often overlook the interdisciplinary complexity and real-world deployment challenges inherent to edge AI systems. In this paper, we present an integrated educational framework, which is designed to bridge the gap by creating a series of modular, hands-on labs and case studies that incorporate practical interactions between hardware and software. Our materials focus on three key aspects of physical edge AI systems: real-time heterogeneous sensing data collection, on-device AI model deployment, and the synergy between edge-based and cloud-based training. Designed for broad adoption for laboratory-level education and seamless integration into existing courses, our materials empower students with practical expertise and interdisciplinary knowledge in edge AI, preparing them to meet the evolving demands of the AI and IoT workforce.

Edge computing has enabled users to experience ubiquitous artificial intelligence (AI) through distributed learning and inference. Continuous efforts to reduce computing burden from the edge devices and increased preservation of privacy have popularized remote inference and federated learning (FL). However, network-level side-channel information can still expose sensitive operational states. In this paper, we demonstrate that network-level telemetry data can be used to fingerprint edge-assisted learning and inference workflows and reveal their operational phases. We propose a hierarchical classification framework, where the first stage separates learning from inference, and the second distinguishes learning phases. In addition, we develop a testbed with convolutional and recurrent neural network-based FL and remote inference systems, alongside an attacker device collecting network sniffing data. Using features derived from flow volumes, transfer speeds, ratios, and latency, the system achieves fingerprinting accuracy of 100% between learning and inference tasks, and 95.9% across different learning phases. These results highlight the vulnerability of edge-assisted distributed AI systems to network-based side-channel fingerprinting.

While modern Autonomous Vehicle (AV) systems can develop reliable driving policies under regular traffic conditions, they frequently struggle with safety-critical traffic scenarios. This difficulty primarily arises from the rarity of such scenarios in driving datasets and the complexities associated with predictive modeling of multiple vehicles. Effectively simulating safety-critical traffic situations is therefore a crucial challenge. In this paper, we introduce TrafficGamer, which facilitates game-theoretic traffic simulation by viewing common road driving as a multi-agent game. When we evaluate the empirical performance across various real-world datasets, TrafficGamer ensures both the fidelity, exploitability, and diversity of the simulated scenarios, guaranteeing that they not only statically align with real-world traffic distribution but also efficiently capture equilibria for representing safety-critical scenarios involving multiple agents compared with other methods. Additionally, the results demonstrate that TrafficGamer provides highly flexible simulations across various contexts. Specifically, we demonstrate that the generated scenarios can dynamically adapt to equilibria of varying tightness by configuring risk-sensitive constraints during optimization. We have provided a demo webpage at: https://anonymous.4open.science/api/repo/trafficgamer-demo-1EE0/file/index.html.

Federated Edge Learning (FEL) enables privacy-preserving, on-device training across heterogeneous edge devices, reducing data transfer costs. However, most FEL approaches remain simulation-based and fail to address realistic on-device training asynchrony caused by variations in computing power, data volume, and energy availability among devices. To address the issue, we propose SPFERE, a Semi-synchronous Power-aware and FairnEss-Regulated Engine in this paper, designed for power-constrained edge environments and implemented on a real-world edge testbed to support asynchronous model updating, power management, and fairness-aware model aggregation. Specifically, we propose a client grouping-based semi-synchronous aggregation protocol that reduces idle waiting time for power-abundant devices and mitigates stale updates from power-constrained devices, along with our in-depth convergence analysis. Then, we introduce a long short-term memory (LSTM)-based power estimation approach to predict remaining battery voltage for devices with limited communication overhead, enabling early warnings for power dropouts. Lastly, we design fusion-based fairness-aware model aggregation methods to prevent bias by considering device participation frequency and training workload. We systematically validate our framework through experiments on both a simulation platform and a real-device testbed. Our extensive experimental results demonstrate the effectiveness and resilience of SPFERE in dynamic and heterogeneous edge environments.

The integration of Large Language Models (LLMs) into Integrated Development Environments (IDEs) introduces a critical vulnerability to indirect Prompt Injection Attacks (PIAs). State-of-the-art coding models, such as Qwen-2.5-Coder, often embed malicious payloads within project configuration rules, resulting in alarmingly high Attack Success Rates (ASR) that compromise local developer environments. To mitigate this threat, this paper proposes IDE-Sanitizer, a preventive, dualmodel defense framework that establishes a zero-trust execution boundary. By combining an HMAC-SHA256 cryptographic state verifier with an air-gapped semantic gate (Llama-3-8B), IDESanitizer decouples intent classification from code generation, preventing attackers from overriding system guardrails. Extensive evaluations across diverse programming frameworks demonstrate that our approach achieves strong robustness, which reduces the ASR to near 0.0% against severe sabotage and exfiltration vectors, while maintaining a near 0.0% False Positive Rate (FPR) on benign workflows. Furthermore, by preemptively blocking malicious payloads before they reach the core generator, IDE-Sanitizer avoids computationally expensive inference loops, reducing average generation latency on adversarial inputs by 89.6%. Ultimately, this architecture offers a secure, efficient, and scalable solution for safeguarding LLM-assisted development cycles.

This article explores the use of reconfigurable intelligent surface (RIS) in the integrated terrestrial-space network (ITSN) to mitigate co-channel interference between low-Earth orbit (LEO) satellite systems and terrestrial base stations (BSs), with a focus on minimizing total power consumption. An alternating optimization framework is developed, decomposing the problem into three subproblems: 1) BS beamforming design solved via second-order cone programming (SOCP) with interior-point methods or zero-forcing (ZF) method with water-filling algorithm; 2) RIS phase-shift optimization addressed through semidefinite relaxation (SDR) combined with Gaussian randomization to handle nonconvex unit-modulus constraints; and 3) LEO satellite transmit signal power optimization addressed by water-filling algorithm. Simulation results show that compared to conventional non-RIS architectures, the proposed system achieves significant power reductions: 14-16 dB for BS and 8.68 dB for satellites. Compared to the weighted minimum mean-square error-based method, it reduces interference at BS users by 4-6 dB and at satellite user by 15-20 dB. When the satellite power is fixed, it still reduces the power consumption of the BS by 6-8 dB. The proposed methodology enables harmonious coexistence between terrestrial BSs and LEO satellite service zones under strict power budgets through enhanced interference decoupling capabilities.

Taking advantage of IoT technologies has enabled the development of more advanced systems across a variety of application domains, including transportation, energy, healthcare, manufacturing, and agriculture [...]

Large Language Models (LLMs) are increasingly used in multi-agent systems, where autonomous agents collaborate on complex tasks through distributed reasoning and communication. While existing work focuses on privacy in single-agent or API-bound settings, inter-agent communication introduces new privacy leakage pathways that remain underexplored. This paper proposes a unified framework for quantifying and mitigating privacy leakage in multi-agent LLMs. It unifies Mutual Information (MI), Bayesian Inference Gain (BIG), and Renyi Differential Privacy (RDP) to jointly quantify empirical leakage, belief refinement, and worst-case privacy loss in multi-agent LLMs. MI captures statistical dependence between private inputs and observable messages, detecting unintended memorization or semantic correlation. BIG models adversarial belief refinement and complements MI by offering a posteriorcentric interpretation of leakage. RDP provides formal privacy guarantees across multi-turn dialogues, quantifying divergence under repeated queries and enabling DP-style defense via output perturbation. We simulate LLM agents with distinct prompts and introduce probing adversaries capable of membership inference, prompt injection, and extraction attacks. The empirical results show that the proposed combined defense reduces MI from 0.35 to 0.10 and the attack success rate (ASR) from 82% to 29%, with minimal Bilingual Evaluation Understudy (BLEU) drop. The combined defense achieves strong privacy with acceptable utility. These findings highlight the importance of privacy-aware agent design and protocol-level defense under black-box threats in multi-agent LLMs.