Telemedicine is becoming a key service domain for future 6G systems, spanning remote consultation, continuous patient monitoring, emergency response, and distributed diagnosis. As these workflows become more mobile, data-intensive, and software-driven, manual service management is increasingly difficult to sustain. Zero-touch management offers a promising path to autonomous operation, but healthcare environments require more than generic automation because adaptation decisions must also reflect clinical priority, trustworthiness, and security risk. This article presents a trust-aware secure zero-touch Internet of Things (IoT) architecture for 6G-enabled telemedicine systems, where Internet of Medical Things (IoMT) devices form the medical sensing and device layer, and telemedicine is treated as a clinically demanding service subset of broader 6G-enabled e-health infrastructures. The proposed architecture combines multi-source observation, continuous trust reassessment, policy-bounded orchestration, and programmable enforcement across device, edge, network, and cloud domains. Structurally, it implements a MAPE-K-aligned closed-loop control cycle integrating a Trust Engine, a Policy Engine with tiered conflict resolution, and a Zero-Touch Orchestrator informed by lightweight AI/ML components. It supports trusted admission and onboarding, anomaly-aware adaptation, selective service degradation, trusted fallback, and auditable self-healing. A remote emergency ambulance scenario demonstrates the step-by-step closed-loop operation, showing how critical medical flows can be preserved while less essential services are adaptively downgraded under network degradation or cyber risk. The article concludes with comparative insights and key open challenges for deploying trustworthy autonomous telemedicine at scale.

A primary challenge in network defense is to mine potential attack campaigns from massive, continuously arriving alerts and telemetry data in real time. This paper presents HoSig-Align I, an edge side unsupervised method designed for network streams to discover homologous blocks. Our approach innovatively fuses heterogeneous features like Internet Protocol (IP) and Payload into a unified representation while strictly excluding temporal information from it, only incorporating time via a dual time scale decay model during graph construction to capture temporal proximity. A density robust similarity is computed using an isolation style random partition forest, leading to a sparse k-Nearest Neighbors (k-NN) graph. The stream is then accurately segmented into internally cohesive and mutually isolated homologous blocks through spectral ordering and contrastive change point detection. Each block is encoded into a lightweight HoSig signature, forming the basis for cross organizational collaboration. Experiments on real network streams show that HoSig-Align I identifies coherent attack campaign blocks and improves separation and boundary clarity over baselines, while meeting low-latency and low-overhead requirements for edge processing.

Cyberbullying and hate speech increasingly appear in multi-modal social media posts, where images and text are combined in diverse and fast changing ways across platforms. These posts differ in style, vocabulary and layout, and labeled data are sparse and noisy, which makes it difficult to train detectors that are both reliable and deployable under tight computational budgets. Many high performing systems rely on large vision language backbones, full parameter fine tuning, online retrieval or model ensembles, which raises training and inference costs. We present a parameter efficient cross-platform multi-modal transfer learning framework for cyberbullying and hateful content detection. Our framework has three components. First, we perform domain adaptive pretraining of a compact ViLT backbone on in domain image-text corpora. Second, we apply parameter efficient fine tuning that updates only bias terms, a small subset of LayerNorm parameters and the classification head, leaving the inference computation graph unchanged. Third, we use noise aware knowledge distillation from a stronger teacher built from pretrained text and CLIP based image-text encoders, where only high confidence, temperature scaled predictions are used as soft labels during training, and teacher models and any retrieval components are used only offline. We evaluate primarily on Hateful Memes and use IMDB as an auxiliary text only benchmark to show that the deployment aware PEFT + offline-KD recipe can still be applied when other modalities are unavailable. On Hateful Memes, our student updates only 0.11% of parameters and retain about 96% of the AUROC of full fine-tuning.

Edge Computing (EC) enables deep neural network training on distributed data, yet it raises significant privacy concerns, particularly under regulations enforcing the “right to be forgotten”. Federated Unlearning (FU) offers a solution by allowing targeted data unlearning without the need for retraining. In Service-Oriented Computing (SOC) systems, where services are composed dynamically and data flows across multiple decentralized nodes, deploying FU introduces additional challenges. Specifically, the lack of direct access to raw data within loosely coupled services, along with the high communication cost required for coordination among distributed components, significantly hinders effective unlearning. Therefore, we propose D3FU, an efficient service-compatible framework that leverages data-free knowledge distillation to achieve self-contained FU. This framework employs local unlearning through Projected Gradient Descent (PGD), which may initially degrade model performance. To mitigate the resulting bias, we integrate Model-Agnostic Meta-Learning (MAML) techniques to generate task-relevant pseudo-samples, thereby enabling data-free distillation and correcting the gradient updates of the local unlearned model. This process effectively restores model performance while ensuring accurate unlearning. Our experimental results, including evaluations of backdoor attacks, demonstrate that D3FU achieves unlearning effects comparable to retraining from scratch, with a maximum reduction in communication cost by up to 32 times.

Security Operations Centers (SOCs) repeatedly revisit the same IP addresses across alerts, logs, tickets, and threat-intelligence updates, yet most LLM-assisted triage systems remain ticket-centric and re-reason over each event in isolation. This design repeatedly consumes overlapping histories, obscures policy-to-action mapping, and weakens cross-visit auditability and runtime robustness. To address this gap, we formulate recurring-IP triage as a governed service problem and present PACT-IP, a policy-governed and auditable continuous IP triage service for SOCs. PACT-IP maintains a bounded working profile plus retained judgment memory for each recurring IP, applies a policy-conditioned significance gate to decide whether newly arrived evidence warrants re-triage, persists structured audit artifacts that retain policy version and decision path, and uses QoSaware mode selection to switch among normal, reuse, cached, fallback, and policy-only execution paths under service pressure. Rather than re-analyzing the full history on every revisit, the service reuses prior structured judgments when updates are not decision-significant and re-triages only when the active policy and new evidence justify it. We evaluate PACT-IP on replayable workloads derived from CTU-13 and CICIDS2017. Against workloadcompatible LLM baselines, PACT-IP improves macro-F1 from 0.842 to 0.869 and AUROC from 0.926 to 0.941 in the strongest head-to-head comparison, while maintaining 99.4% policy compliance, 98.2% audit completeness, 94.5% revisit consistency, and 2.14 s $p 95$ latency. The gains are moderate on predictive quality but larger on governance and runtime robustness, suggesting that recurring-IP triage is better handled as a governed service than as isolated ticket-level reasoning. The implementation is available at https://github.com/shaoandy1211/PACT-IP.

Synanthropic flies are major disease vectors in China, yet routine identification still relies on bulky Chinese-language keys and monographs that are hard to use, slow to update, and inaccessible to many practitioners. We propose a locally deployed, server-side retrieval-augmented generation (RAG) platform that transforms static entomological monographs into an interactive identification and learning environment. Authoritative books are ingested on an on-premise server and accessed through a Chinese conversational interface supporting forward (key-driven) and reverse (trait-driven) identification, plus explanatory queries on taxa and morphology. Using the RAG-Anything toolkit, we build a domain knowledge graph and couple it with dense retrieval and a multimodal large language model (Qwen3-VL-32B-Instruct) served via vLLM. The system currently covers 1,671 vector species with complete family-, genus-, and species-level keys and runs fully on local GPUs, preserving data sovereignty while modernizing vector surveillance workflows. The GitHub repository is available here. 1

Subspace clustering is one of the most popular clustering methods due to its effectiveness. Although subspace clustering methods have been demonstrated to achieve promising performance, they still lack interpretability, especially when handling high-dimensional complicated data. To bridge this gap, this paper focuses on the interpretability of subspace clustering and proposes a novel interpretable subspace clustering method. Our goal is to answer two key questions about the interpretability in subspace clustering: (1) when handling an individual sample, which features should work for this sample? (2) Which cluster or subspace will the features that work put this sample into? To answer these two questions, we design two new interpretability regularized terms and plug them into the subspace clustering. In this way, we show that interpretability can be used to improve the clustering performance in turn. Extensive experiments on benchmark data sets demonstrate the effectiveness of our method in terms of clustering performance and interpretability.

Last mile delivery by drones is a core component of innovative logistics systems, relying heavily on AI models for essential operations such as path planning and object recognition. However, models trained on region specific data often experience significant performance degradation when deployed in unfamiliar environments due to geographic domain shifts. This limitation impedes the rapid deployment of logistics networks and hinders model adaptation. Federated Learning (FL), as a distributed machine learning paradigm, enables multiple clients with diverse data to collaborate in training a global model. Nevertheless, within Mobile Edge Computing (MEC) environments, FL faces critical challenges, including data bias, high drone mobility, and intermittent communication windows between drones and edge servers. This paper proposes FedMO, a mobility aware FL framework for drone based last mile delivery with edge cloud collaboration. FedMO introduces a novel algorithmic insight by treating the drone's flight path as a unified proxy for both communication reliability and data distribution heterogeneity. The framework implements a synergistic three stage selection policy that jointly optimizes connectivity success, data value, and resource efficiency. This effectively transforms mobility from a disruption risk into a diversity enhancing asset. Experimental results using real world drone video datasets demonstrate that FedMO improves convergence speed by approximately 15% compared to baseline methods with only $\mathbf{3 0 \% - 4 0 \%}$ client selection. With equivalent client participation, FedMO achieves a 25% improvement in convergence speed over the FedAvg algorithm.

As facial recognition becomes increasingly integrated into consumer Internet of Things (CIoT) ecosystems such as smart cameras, mobile devices, and home surveillance protecting multimedia identity data while retaining utility and ensuring security has become a pressing challenge. Existing anonymization techniques often result in irreversible transformations that prevent legitimate identity recovery, limiting their applicability in scenarios like access control or forensic verification. To address this, we propose a reversible facial anonymization framework designed for secure multimedia processing in CIoT environments. Our approach combines Reversible Noise Injection (RNI) for learnable encryption, Hybrid Adversarial Training (HAT) for privacy preserving transformation, and a Zero Trust Identity Recovery (ZTIR) module that enables authorized identity restoration through cryptographic key verification. The system enforces security through multi factor authentication, TLS encrypted communication, and optional blockchain based key management. Implemented on edge devices, the framework supports real time anonymizationwith low computational overhead and empirically strong privacy protection, as measured by reduced recognition accuracy and perceptual or distributional metrics. These results validate the framework’s suitability for privacy preserving and secure multimedia intelligence in real world CIoT deployments.

With the increasingly severe trend of social aging, the detection measures for preventing the elderly from falling have become more and more urgent. And there are also more and more detection methods for the elderly with the development of deep learning, and the accuracy is getting higher and higher. Against this backdrop, we propose an innovative architecture combined with an experimental procedure suitable for the gait of the elderly, it classifies and detects the abnormal gait of the elderly based on the data obtained from the IMU (Inertial Measurement Unit) and Kinect. This architecture can integrate the information from the two experimental devices, addressing the problems of insufficient classification accuracy and inadequate information in the past single-modal gait information. Through ablation experiments, we have determined that the multi-modal gait anomaly detection method based on our architecture is more effective than the single-modal detection method. At the same time, when using the same multi-modal data, our architecture has made significant improvements compared with the traditional methods. In addition, this architecture can be adapted to the data from devices other than the IMU and Kinect. Before using it, it only needs to align the sequential data in time. query

Artificial Intelligence of Things (AIoT) systems are widely utilized in various domains such as smart logistics and smart health. In AIoT systems, computational tasks earmarked for offloading by IoT devices often have different task priorities and heterogeneous resource requirements. However, these factors will affect the QoS of the solution and the task response time. Therefore, motivated by a typical smart warehouse system, a Multi-access Edge Computing (MEC) based distributed task offloading framework is proposed. Afterward, a delaysensitive mixed-integer programming model is defined to address the imbalanced utilization of computing resources between edge and cloud servers during the offloading process. Finally, based on the attention mechanism, a Priority-Driven Multi-Agent Deep Reinforcement Learning algorithm (PDMA-DRL) is proposed to fulfill real-time resource requirements while reducing the task response time. The PDMA-DRL algorithm consists of two agents, and each agent corresponds to a sub-problem. Comprehensive experimental results demonstrate that our proposed PDMA-DRL algorithm can outperform representative existing methods in generating delay-sensitive task offloading plans and effectively reducing the task response time.

Low-altitude unmanned aerial vehicle (UAV) vision is critically hindered by the Sim-to-Real Gap, where models trained exclusively on simulation data degrade under real-world variations in lighting, texture, and weather. To address this problem, we propose RA3T (Region-Aligned 3D Transformer), a novel self-supervised framework that enables robust Sim-to-Real adaptation. Specifically, we first develop a dual-branch strategy for self-supervised feature learning, integrating Masked Autoencoders and contrastive learning. This approach extracts domain-invariant representations from unlabeled simulated imagery to enhance robustness against occlusion while reducing annotation dependency. Leveraging these learned features, we then introduce a 3D Transformer fusion module that unifies multi-view RGB and LiDAR point clouds through cross-modal attention. By explicitly modeling spatial layouts and height differentials, this component significantly improves recognition of small and occluded targets in complex low-altitude environments. To address persistent fine-grained domain shifts, we finally design region-level adversarial calibration that deploys local discriminators on partitioned feature maps. This mechanism directly aligns texture, shadow, and illumination discrepancies which challenge conventional global alignment methods. Extensive experiments on UAV benchmarks VisDrone and DOTA demonstrate the effectiveness of RA3T. The framework achieves +5.1% mAP on VisDrone and +7.4% mAP on DOTA over the 2D adversarial baseline, particularly on small objects and sparse occlusions, while maintaining real-time performance of 17 FPS at 1024 × 1024 resolution on an RTX 4080 GPU. Visual analysis confirms that the synergistic integration of 3D geometric encoding and local adversarial alignment effectively mitigates domain gaps caused by uneven illumination and perspective variations, establishing an efficient pathway for simulation-to-reality UAV perception.

In recent years, emerging technologies like UAV (Unmanned Aerial Vehicle) delivery, Metaverse, and blockchain become more popular, making big changes in many areas of work and life. In this paper, we propose a lightweight blockchain-based protocol designed to facilitate UAV delivery services within the metaverse, thereby enhancing the efficiency and reliability of metaverse services. Given the interest in the metaverse as a platform for immersive virtual experiences, integrating UAV delivery presents novel challenges and opportunities. This paper aims to bridge the gap between virtual and physical service delivery, leveraging the decentralized, secure, and transparent nature of blockchain technology. The lightweight protocol not only ensures secure transactions and verifiable service fulfilment within the metaverse but also enables a seamless interface between virtual service orders and their physical execution via UAVs. Through a series of simulations, we demonstrate the feasibility, scalability, and security of our protocol, highlighting significant improvements in service delivery time, reduction in fraudulent activities, and enhanced user trust.

With the rapid expansion of Artificial Intelligence (AI) services, smart devices generate a large amount of user data at the edge network, which urgently needs to be protected while effectively extracting information. Federated learning (FL) is an important technology for handling dispersed data and strict privacy requirements in this context. However, the security threats caused by model inversion attacks and poisoning attacks can affect the mutual trust between the client and server. Yet, for these two types of attacks, the existing defense mechanisms are contradictory in terms of whether the model parameters are publicly disclosed. In addition, the data distribution of the clients is imbalanced which will increase the bias of model, reducing its practicality. To address this issue, this study proposes a dual defense selfbalanced federated learning (DDSFL) framework, aiming to introduce a novel lightweight defense mechanism during the model parameter aggregation stage, combating these two types of attacks simultaneously by applying differential privacy and adjusting learning rates. In addition, this method also integrates a middleware-based reordering algorithm to enhance the robustness of the framework. Experimental results show that DDSFL effectively improves the ability to resist imbalanced data, forged data, and malicious behavior, significantly enhancing the generalization performance and security of the FL system.

Federated learning allows multiple parties to collaboratively train AI models without sharing raw data—but this privacy advantage comes with risks. A major threat is poisoning attacks, where malicious participants submit fake model updates to sabotage the global model. Current defenses typically use statistical methods to filter out suspicious updates, but these approaches struggle to detect subtle attacks in real time and offer little insight into why an update was flagged. To address this, we introduce Incremental Provenance Analysis (IPA), a new defense framework that monitors how model updates evolve over time W . Unlike traditional methods, IPA doesn’t just look for outliers—it learns the typical “provenance” of benign updates. When an update deviates suspiciously, IPA not only detects it but also traces the attack source and adjusts the aggregation strategy dynamically. Our experiments show IPA effectively thwarts poisoning attempts while maintaining model accuracy, offering a more transparent and adaptive solution for secure federated learning.

Unmanned Aerial Vehicles (UAVs), play a significant role in the advancement of pervasive systems by providing efficient, scalable, and innovative solutions in various sectors, such as smart cities or location-based services. However, the current UAV delivery scenario presents various challenges for recipients, including lengthy identity verification processes, privacy concerns, and risks of fraud and theft. In response to these issues, this paper proposes an innovative system that leverages Blockchain technology and Adversarial Machine Learning (AML) to tackle these problems effectively. The proposed system streamlines the verification process, enhances privacy safeguards, and reduces fraud risks. The integration of AML is crucial as it enables users to have greater control over their personal data, boosting privacy and security. AML also plays a critical role in this system by creating test scenarios that reinforce the machine learning model against adversarial threats, ensuring precision and dependability in the face of malicious manipulations. The paper also provides details on the practical implementation and evaluation of this system in real-life adversarial situations. The evaluation results demonstrate superior performance on selected metrics, highlighting the potential of this system as an effective solution for verifying recipients in UAV delivery.

In recent years, there has been a rapid increase in the integration of Internet of Things (IoT) systems into edge computing. This integration offers several advantages over traditional cloud computing, including lower latency and reduced network traffic. In addition, edge computing facilitates the protection of users’ sensitive data by processing it at the edge before transmitting it to the cloud using techniques such as federated learning (FL) and differential privacy (DP). However, these techniques have limitations, such as the risk of user information being obtained by attackers through the uploaded weights/model parameters in FL and the randomness of DP, which limits data availability. To address these issues, this paper proposes a framework called FedShufde (Federated Learning with a Shuffle Model and Differential Privacy in Edge Computing Environments) to protect user privacy in edge computing-based IoT systems, using an unmanned aerial vehicle (UAV) delivery system as an example. FedShufde uses local differential privacy and the shuffle model to prevent attackers from inferring user privacy from information such as UAV’s location, flight conditions, or delivery address. In addition, the network connection between the UAV and the edge server cannot be obtained by the cloud aggregator, and the shuffle model reduces the communication cost between the edge server and the cloud aggregator. Our experiments on a real-world edge-based smart UAV delivery system using public datasets demonstrate the significant advantages of our proposed framework over baseline strategies.

In edge intelligence scenarios, Federated Learning (FL) has emerged as a preeminent method for ensuring privacy preservation and fostering collaboration among end devices. This is primarily due to its ability to train a global model in an edge computing environment without compromising the data privacy of end devices. However, challenges arise in the form of non-independent and identically distributed (Non-IID) issues, such as heterogeneous graph data distributions, stemming from heterogeneity in end device data. These issues notably impede the convergence speed of the federated model and escalate communication costs. To mitigate these issues, we introduce a novel edge computing based FL framework with a Dual Attention mechanism, termed FedDAD. Specifically, FedDAD integrates a Graph Attention Network (GAT) within the local model training process, which is designed to capture the variety of relationships across diverse client data distributions. Concurrently, FedDAD employs an attention mechanism across different client models to fuse key features from various client models into the global model. The efficacy of FedDAD is evaluated through extensive experiments on three heterogeneous graph datasets, demonstrating its superior performance over representative FL baselines.

Edge-Integrated Unmanned Aerial Vehicles (UAVs) delivery systems have demonstrated the advantage of higher efficiency and lower latency in comparison with traditional intelligent delivery systems. But with its rapid development, a series of security and privacy issues have also emerged. For instance, it is of vital importance to maintain data safety due to UAVs exchanging sensitive data with servers through public channels, attackers can easily gain access to sensitive information by launching attacks including man-in-the-middle and impersonation attacks. Additionally, the requirements of frequent authentications between UAVs and edge servers can result in increased computation overhead, while UAVs are fast-moving and resource-constrained, and excessive computational overhead can degrade the user experience. To address these challenges, this paper proposes a Two-Stage Biometric-Assisted Authentication Framework (TBAF) that enhances security and efficiency. In TBAF, a novel secret sharing method is designed to distribute storage biometric templates with protection, ensuring the secret values which are biometric templates can only be accessed by authorized parties. Additionally, the two-stage authentication protocol reduces computation and communication overhead. Extensive formal and informal security analysis confirms the superior performance of the proposed protocol compared to existing solutions.