The burgeoning 5G network brings edge servers closer to users to host online applications. These edge servers are typically kept running 24/7 to meet users' computational demands. However, the user coverage, privacy assurance, and service delay have consistently undermined users' confidence, compounded by the significant environmental damage caused by excessive energy consumption. Recently, various approaches have been proposed to tackle the energy-saving demand response issue in the multi-access edge computing (MEC) system. Unfortunately, existing attempts often compromise service quality and energy efficiency for privacy enhancement, and incur significant computational overheads and delays unsuitable for real-time services. Therefore, maintaining satisfying user coverage with energy consumption while adhering to users' privacy demands with low computational overhead is critical to achieving sustainable edge services. To address those challenges, we systematically formulate the location-privacy-preserving edge demand response (LEDR) problem and introduce a novel system named KGEES. KGEES incorporates $k$k-anonymity geo-obfuscation to enhance user privacy while leveraging a heuristic approach to finalize resource allocation strategies under geo-distortion greedily to jointly improve system utility, energy, and time efficiency. Comprehensive experiments on a real-world dataset demonstrate that KGEES surpasses the representative approaches by an average of 1.187 & times; in system utility and 1.192 & times; in energy efficiency while being 203.5 & times; faster.

Next networks demand federated learning (FL) frameworks that ensure both privacy and robustness under heterogeneous non-IID data. Traditional FL approaches, such as FedAvg, suffer significant accuracy degradation in non-IID settings, while naive differential privacy (DP) mechanisms reduce model utility. To address this, we propose Differentially Private Model Recombination as a Service (DP-MRaaS), which integrates layer-wise model recombination, adaptive DP, and a Thompson sampling controller. DP-MRaaS dynamically alternates between aggregation and recombination while allocating noise and clipping based on loss improvements. The service abstraction exposes these decisions as controllable management functions for privacy-aware model training in network and service management systems. Theoretical analysis confirms convergence rates comparable to FedAvg, and extensive experiments on CIFAR-10 and Fashion-MNIST demonstrate its effectiveness. On CIFAR-10 CNN, DP-MRaaS achieves 57.78 +/- 0.51% accuracy under IID, matching FedAvg (57.78%) while preserving (is an element of, delta) = (3.5, 10(-5)) privacy. On CIFAR-10 ResNet-18, DP-MRaaS attains 66.57 +/- 0.54% (IID) and 61.30 +/- 0.30% (Balanced Non-IID), competitive with FedAvg-DP (66.70%, 61.84%) and Ditto-DP (66.86%, 61.76%). On Fashion-MNIST, DP-MRaaS achieves 90.52 +/- 0.10% (IID) and 90.02 +/- 0.07% (Balanced Non-IID), the highest among DP-enabled methods. A five-way ablation study confirms that recombination is the primary accuracy driver (+8.6% on ResNet-18), while adaptive DP preserves utility with < 0.1% degradation on CNN. These results show that DP-MRaaS achieves a superior balance between accuracy, stability, and privacy, advancing the trustability of FL for service management in next-generation networks.

AI-powered face recognition has become essential to various IoT applications, including home automation, security systems, and personalized services. While these systems offer significant advancements, they still face critical challenges related to accuracy and privacy. One major issue is class imbalance, which is common in face recognition systems where certain demographic groups are underrepresented. This imbalance results in biased models, compromising the accuracy and fairness of these systems. Furthermore, traditional centralized training methods can expose sensitive facial data, raising serious privacy concerns. Federated Learning (FL) has emerged as a solution to improve model training by enabling collaboration across devices without sharing sensitive data. However, it also worsens the issue of data heterogeneity. This paper proposes a Hierarchical Federated Learning (HFL) framework to address class imbalance while preserving privacy. By aggregating local models at different hierarchical levels, the framework mitigates data imbalance and enhances fairness in face recognition systems. Additionally, a privacy-preserving mechanism based on Secure Multi-Party Computation (SMPC) is implemented to ensure data security during the training process.

Federated learning (FL) enables collaborative biometric model training without centralizing raw face images, but it does not prevent privacy leakage through trained models, output scores, embeddings, or shared updates. We propose an input-level privacy pipeline for federated face recognition that combines ρ–n CutMix, keyed block-wise permutation, and a parameter-free differentiable inverse decoder. On each client, multi-source mixing reduces direct attribution to a single identity, keyed permutation disrupts canonical spatial structure, and inverse decoding restores the block order required for authorized local learning. The contribution is this role-separated client–model interface, which operates without changing the recognition backbone, optimizer, FedAvg aggregation rule, or model-update format. We evaluate the pipeline on three face datasets—PubFig, LFW, and PINS—and use CIFAR-10 as a non-biometric control, with ResNet-18 and ResNet-34 backbones. The evaluation covers classification utility, biometric verification, score-based membership inference, calibration, implementation-level latency, and adaptive white-box model inversion. The inversion attacker jointly optimizes a synthetic image and a Sinkhorn-relaxed permutation proxy without access to the secret key, using multiple random restarts. The results demonstrate dataset-, backbone-, and configuration-dependent privacy–utility trade-offs. Several defended configurations reduce membership separability or weaken structural and identity-level reconstruction relative to the raw-input baseline, although the improvements are not uniform across all metrics. The proposed pipeline therefore provides an empirical input-level defense that can complement formal and system-level privacy mechanisms rather than replace them.

In the domain of computation offloading from mobile devices, Space-Air-Ground Integrated Networks ) emerge as a potent paradigm, leveraging unmanned aerial vehicles (UAVs) and low earth orbit (LEO) satellites as computational resource providers. One prevalent strategy to reducing energy consumption within SAGIN systems is to jointly optimize UAV deployment and computation offloading. However, given the variable nature of mobile device positioning and computational requisites, the deployment and offloading must be continually recalibrated. This joint optimization faces a significant challenge of combinatorial explosion. Addressing the difficulties of real-time computation tasks necessitates efficacious methodologies for devising joint optimization schemes. In this paper, we introduce a convex optimization-based algorithm to minimzie the weighted total energy consumption within the SAGIN framework by jointly optimizing UAV deployment and computation offloading. The algorithm decomposes the original problem into a set of sub-problems: UAV deployment, ground device (GD) access, and computation offloading, employing the Block Coordinate Descent (BCD) method. By alternately addressing these sub-problems, the algorithm derives a near-optimal solution efficiently. Simulation results demonstrate that our approach can generate a joint optimization solution in a few seconds and diminish the weighted total energy consumption than other classic methods by 1.5% similar to 10.97%.

The proliferation of Internet of Things (IoT) devices has intensified the need for scalable and energy-efficient federated learning (FL). While hierarchical federated learning (HFL) improves scalability by adding an edge aggregation tier, it still suffers from high communication costs, slow convergence, and degraded accuracy under non-independent and identically distributed (non-IID) data. Existing methods, such as quantization, sparsification, and static pruning, alleviate specific bottlenecks but fail to jointly optimize efficiency, robustness, and accuracy. This article proposes an intent-driven dual-layer model pruning framework for HFL, where an energy management system (EMS) and an Intent-driven Dual-layer Pruning Orchestrator (IDPO) dynamically translate system-level intents (e.g., energy minimization or accuracy preservation) into pruning actions at both edge and cloud layers. Experiments on MNIST, CIFAR-10, and FEMNIST show up to 41% smaller models, 12 & times; faster training, 28%-35% lower energy use, and +12.9% accuracy gain under non-IID data, establishing the framework as a robust and sustainable solution for IoT learning

Wearable devices are increasingly utilized in healthcare to support continuous monitoring, diagnosis, and personalized intervention across a wide range of medical conditions. Despite their growing adoption, healthcare practitioners often lack the ability to inspect, interpret, or adapt the artificial intelligence models deployed on these devices. Many such models operate as opaque systems, while others require substantial technical expertise to understand, limiting their clinical customization. Moreover, complex machine learning models are frequently too computationally demanding for resource-constrained wearable platforms. To address these challenges, this paper proposes HealthwearSPMG, a semantic-driven framework that leverages large language model based agents to extract critical clinical decision logic from both healthcare practitioners and large-scale machine learning models, and subsequently distills this knowledge into compact, deployable models optimized for specified wearable hardware configurations. The proposed approach reduces the need for costly retraining while enabling dynamic generation of lightweight, task-specific models tailored to diverse medical use cases. Our extensive experimental results on three diverse medical datasets show that the proposed framework improves F1-scores by 0.99%, 1.46%, and 4.69%, while using less than 0.025% of the context window tokens, demonstrating that the framework achieves strong predictive performance, interpretability, and high LLM token-efficiency.

Medical institutions face significant challenges in collaborative machine learning due to diverse feature distributions, strict privacy regulations, and the critical issue of data scarcity. Traditional federated learning approaches often struggle with poor generalization and inadequate privacy protection in these sensitive healthcare settings. This paper presents a Quantum-Enhanced Vertical Federated Learning (Q-VFL) framework that effectively combines quantum feature encoding with supervised contrastive learning. Our approach employs a three-stage quantum circuit architecture for robust feature encoding, followed by classical neural networks, implemented through a privacy-preserving split learning procedure. Experimental results across diverse medical imaging datasets demonstrate that the proposed Q-VFL framework achieves strong performance, with accuracy gains of up to 25 percentage points over quantum cross-entropy baselines and competitive results against classical methods. Beyond raw performance, the framework exhibits high clinical reliability and stability. Crucially, it provides stronger privacy guarantees: detailed analysis shows that Q-VFL reduces membership inference attack success rates by 10–15

In next-generation 6G Internet of Things (IoT) networks, semantic communication (SemCom) has emerged as a key paradigm that transforms raw data into high-level feature representations, thereby reducing communication overhead while enhancing interpretability. When combined with federated learning (FL), these semantic embeddings enable decentralized model training without centralizing raw data, preserving user privacy, and supporting large-scale collaboration. However, semantic features may inadvertently encode sensitive information or act as adversarial triggers, introducing new privacy risks that current unlearning techniques fail to address. To overcome this challenge, we propose selective semantic feature unlearning (SSFU), a novel framework that performs unlearning at the feature level rather than at the client level. SSFU employs an ensemble-based risk scoring mechanism to identify high-risk latent components, followed by gradient ascent and semantic masking to remove their influence. Unlike existing methods that depend on costly retraining or full client exclusion, SSFU preserves benign semantic knowledge and allows training to continue with minimal disruption. The framework guarantees bounded convergence, and empirical results on benchmark datasets show that SSFU effectively eliminates sensitive features while maintaining predictive accuracy. SSFU thus represents a robust, privacy-preserving FL framework tailored for SemCom in 6G IoT systems.

Edge-enabled sensing applications in digital health, industrial Internet of Things (IIoT), and extended reality (XR) face a persistent trilemma: lightweight edge models meet strict latency, bandwidth, and energy budgets but degrade under noise and distribution shift; cloud models improve robustness and calibration but can violate tail-latency and privacy constraints; and na & iuml;ve split inference reduces edge compute at the cost of inflated network usage and unstable tails. We present a deployable, edge-first blueprint for budget-aware collaborative training and co-inference between large cloud models and lightweight edge models in sensor-cloud networks. The design has three pillars: (i) a collaborative training workflow that transfers robustness and calibrated confidence from a cloud teacher to an edge student so most inference completes locally; (ii) a confidence-gated, budget-aware router that selects among EDGE, EDGE+CORRECTION, and CLOUD actions using calibrated uncertainty and short-horizon link telemetry while explicitly enforcing P95 latency, bandwidth, and energy budgets; and (iii) privacy scaffolding via trusted execution environments (and optional differential privacy), with all correction or offload overheads accounted end-to-end when invoked. Compact case studies on inertial and physiological sensing workloads demonstrate edge-dominant operation with tight latency tails, small and predictable bandwidth, bounded energy consumption, and improved robustness over non-collaborative baselines. We conclude with practical guidance on system design, scheduling, measurement, and deployment trade-offs for real-world sensor-cloud intelligence.

The global rollout of 5G networks has spurred the rapid deployments of edge servers for hosting latency-sensitive web applications, which improves quality of experience (QoE). However, current efforts fall short in the substantial energy costs associated with the 24/7 operation of edge servers and overlook user privacy by requiring accurate user information for service provision, eroding the sustainability of multi-access edge computing (MEC). To enhance the QoE and service performance while ensuring privacy in MEC, we systematically formulate the inter action among edge servers as a privacy-preserving experience aware edge resource control (PEERC) problem. To address this, we conduct a global resource control and propose a collaborative resource allocation system named MERA. MERA leverages k anonymity data obfuscation to protect user location and resource demand privacy while enhancing service performance and energy efficiency with mean-field multi-agent reinforcement learning. Extensive experiments based on a synthetic real-world dataset demonstrate that MERA significantly surpasses benchmarks in terms of QoE, user coverage, privacy, and energy efficiency by 1.18×, 1.24×, 1.63×, and 1.27× on average.

Large Language Models (LLMs) increasingly underpin intelligent web applications, from chatbots to search and recommendation, where efficient specialization is essential. Low-Rank Adaptation (LoRA) enables such adaptation with minimal overhead, while federated LoRA allows web service providers to fine-tune shared models without data sharing. However, in privacy-sensitive deployments, clients inject varying levels of differential privacy (DP) noise, creating privacy heterogeneity that misaligns individual incentives and global performance. In this paper, we propose WinFLoRA, a privacy-heterogeneous federated LoRA that utilizes aggregation weights as incentives with noise awareness. Specifically, the noises from clients are estimated based on the uploaded LoRA adapters. A larger weight indicates greater influence on the global model and better downstream task performance, rewarding lower-noise contributions. By up-weighting low-noise updates, WinFLoRA improves global accuracy while accommodating clients' heterogeneous privacy requirements. Consequently, WinFLoRA aligns heterogeneous client utility in terms of privacy and downstream performance with global model objectives without third-party involvement. Extensive evaluations demonstrate that across multiple LLMs and datasets, WinFLoRA achieves up to 52.58% higher global accuracy and up to 2.56x client utility than state-of-the-art benchmarks. Source code is publicly available at https://github.com/koums24/WinFLoRA.git.

As the 5G network rapidly expands, more edge servers are being deployed to provide more efficient and low-latency mobile services. However, limited edge resources constrain users' demand response, while continuous server operation leads to significant energy consumption, corroding the sustainability of the multi-access edge computing (MEC) system. Existing resource allocation methods depend on accurate user locations, risking privacy exposure, while protection techniques often cause significant service degradation due to spatial distortion. Moreover, user mobility in MEC systems poses new challenges for edge resource management, which requires dynamic server collaboration and user data migration, which incur additional costs and delays. To address these challenges, we propose MoSEEC, which employs user-adaptive differential geo-obfuscation to secure trajectory privacy while dynamically enhancing service performance with energy awareness. Our results demonstrate its superior performance in migration delays and system utility by 1.54 & times; faster and 1.15 & times; higher compared to existing techniques with privacy guarantees, respectively. In addition, our system outperforms state-of-the-art approaches by 5 & times; faster on average in terms of computation overhead.

Edge computing enables distributed intelligence in resource-constrained IoT environments. However, traditional Federated Learning (FL) struggles with heterogeneous device capabilities, dynamic network conditions, and non-IID data distributions, resulting in straggler effects, slow convergence, and inefficient resource utilization. This paper proposes Resource-Aware Dynamic Split Federated Learning (RAD-SFL), a framework for efficient distributed training in heterogeneous edge environments. RAD-SFL introduces a dynamic model layer splitting mechanism that adaptively partitions model execution between client devices and edge servers based on real-time computation and communication conditions, and a group-and-reorder technique that organizes devices into balanced groups with similar data distributions to improve model convergence under non-IID settings. We validate RAD-SFL through experiments on widely adopted datasets using both a simulated environment and a real testbed with heterogeneous IoT devices. Results demonstrate that RAD-SFL reduces the training time by up to 66.4%, decreases device-side energy consumption by 52.5%, and improves global model accuracy by up to 30.5% compared to FL and SFL baselines.

Large language model (LLM)-based multi-agent systems (MAS) have demonstrated exceptional capabilities in solving complex tasks, yet their effectiveness depends heavily on the underlying communication topology that coordinates agent interactions. Within these systems, successful problem-solving often necessitates task-specific group structures to divide and conquer subtasks. However, most existing approaches generate communication topologies in a node-centric manner, leaving group structures to emerge implicitly from local connectivity decisions rather than modeling them explicitly, often leading to suboptimal coordination and unnecessary communication overhead. To address this limitation, we propose GoAgent (Group-of-Agents), a communication topology generation method that explicitly treats collaborative groups as the atomic units of MAS construction. Specifically, GoAgent first enumerates task-relevant candidate groups through an LLM and then autoregressively selects and connects these groups as atomic units to construct the final communication graph, jointly capturing intra-group cohesion and inter-group coordination. To mitigate communication redundancy and noise propagation inherent in expanding topologies, we further introduce a conditional information bottleneck (CIB) objective that compresses inter-group communication, preserving task-relevant signals while filtering out redundant historical noise. Extensive experiments on six benchmarks demonstrate the state-of-the-art performance of GoAgent with 93.84

Network function virtualization enables flexible network services through Service Function Chain (SFC) deployment. Existing SFC deployment solutions rely on mixed integer linear programming (MILP) solvers or heuristics. However, they fail to support multi-objective deployment and adaptive service operation. In this article, we propose a large language model (LLM) and mixture-of-experts (MoE)-enabled framework for multi-objective and adaptive SFC deployment (MoED). In detail, MoED considers a comprehensive set of fundamental optimization objectives and deploys specialized expert agents for them. Then, MoED leverages LLMs to analyze input user intents and formulates the optimization function. Next, it uses LLMs to dynamically activate expert agents and coordinate their output deployment strategies. Experimental results demonstrate that compared to existing solutions, MoED reduces the power consumption by up to 68.4% while achieving near-optimal performance in terms of throughput, latency, and resource consumption , and reduces deployment strategy solving time by several orders of magnitude.

Personalized Federated Learning (pFL) tailors models to each client’s local data distribution in heterogeneous federated learning settings. Federated Meta-Learning (FML) is a branch of pFL that uses meta-learning to achieve fast adaptation, where clients start with a meta-model and personalize it by fine-tuning it with local data. Since a single global meta-model has limitations when the data distribution of clients varies significantly, meta-model personalization should be considered in FML. However, most benchmark pFL methods lack meta-model personalization, and usually lack meta-learning or relying on a single global meta-model. Besides, these methods can neither provide meta-model personalization nor guarantee generalization and convergence, due to the challenges in measuring the distance between the meta-model and the client model in FML. To address these issues, we combine FML with hypernetwork and propose a constrained hypernetwork-based FML framework called FMLH, which innovatively utilizes hypernetwork to capture the differences in fine-tuned models, thereby providing personalized meta-models for each client. We provide rigorous mathematical proofs illustrating how the hypernetwork affects the convergence and generalization bounds of FMLH. Experimental results demonstrate that FMLH significantly improves the generalization of the model in cross-client shifts, with the lowest decile accuracy improved by up to 18.71%. FMLH also outperforms representative pFL algorithms by up to 5.6% in terms of maximum accuracy improvement.

Deep learning models are increasingly deployed in wireless security and privacy domains, particularly on wireless-enabled Internet-of-Things (IoT) and edge platforms. While such deployments enable real-time and scalable intelligence, they also raise significant privacy and security challenges in adversarial wireless environments. These challenges arise when models are exposed to data-poisoning and backdoor attacks through untrusted data sources or compromised IoT nodes. As such, sensitive data must be removed after training. Machine unlearning (MU) provides a principled mechanism to selectively erase the influence of specific data classes from trained models, offering both privacy protection and an effective defense against adversarial manipulation. However, the impact of unlearning on model reasoning and trustworthiness remains insufficiently understood. This paper presents an Explainable Machine Unlearning (XMU) framework that integrates MU with explainable AI (XAI) techniques to analyze how selective forgetting affects deep model behavior in security-sensitive wireless and IoT settings. Using GradientSHAP, XMU visualizes attribution changes of the models before and after unlearning, while an Explanation Divergence Metric (EDM) based on Jensen–Shannon divergence quantitatively measures the shifts in model explanations. Experimental results on edge-representative sensing workloads demonstrate that the proposed XMU framework effectively removes the influence of poisoned or sensitive samples, disperses class-specific activations, and preserves stable explanations for retained data. These findings establish XMU as a privacy-preserving, security-enhancing, and interpretable defense mechanism for deep learning models operating in wireless, IoT, and edge intelligence environments.

Searchable encryption enables privacy-preserving queries over data outsourced to cloud services. Classical symmetric schemes deliver efficient search but largely assume a single client setting; multi-client variants permit delegation yet typically treat authorization as an owner-local decision, overlooking regulations enforced by higher-level authorities (e.g., sector-specific compliance). In practice, limited familiarity with regulatory detail or operational lapses can lead owners to delegate search permissions that violate authority regulations, rendering existing systems unsuited to regulated, multi-client cloud environments. To address this gap, we propose two systems. First, we propose SENTRY, a multi-client dynamic searchable encryption framework with a built-in compliance-check service via sanitized authorization. Its core is a tag-based sanitization protocol: the authority encodes prohibited keywords as hidden tags, and the sanitizer uses these tags to remove non-compliant per-keyword search permissions before they reach readers, without learning the underlying keywords. As a result, readers receive only compliant search capabilities. We then propose F-SENTRY, a forward private variant of SENTRY for settings where regulations evolve over time. F-SENTRY preserves the same sanitized-authorization mechanism and further adds forward privacy through a tailored constrained shiftable encryption, which binds search permissions and newly added encrypted updates to regulatory epochs. Consequently, permissions issued before a policy change cannot be used to retrieve data added afterward unless they are refreshed for the new epoch. We formalize the security of both SENTRY and F-SENTRY and prove them secure under standard assumptions. Experiments under cloud-like workloads show that SENTRY achieves regulation-compliant authorization with modest over head, while F-SENTRY provides stronger protection under changing regulations at practical additional cost.

This study introduces a semi-asynchronous web browser-based federated learning (FL) framework with secure communication, designed for heterogeneous IoT networks. The study addresses the complex challenges of FL over a large number of heterogeneous IoT devices connected by insecure and unstable communication links compounded with the possibility of client dropout, where clients may disconnect due to internal failures or external reasons, using a model-centric client selection strategy and a semi-asynchronous FL protocol. In the semi-asynchronous process, the central server emulates a synchronous system model by setting a submission deadline for clients and only timely updates are considered for the client selection process. This ensures a robust FL process for the client dropout problem. Unlike agnostic selection protocols, we propose a performance-based client selection algorithm that evaluates local model quality. This maintains fairness and global model performance even with intermittent client participation. Moreover, we utilize TLS-based central server validation and a public key cryptography-based client authentication mechanism to defend against communication layer threats (man-in-the-middle attacks, unauthorized client access) to further strengthen communication security. The proposed framework is hereafter referred to as SecHeto-FL. Extensive evaluation results show that SecHeto-FL significantly outperforms state-of-the-art techniques, in terms of global model accuracy. Theoretical analysis further guarantees convergence of the global model and resilience against client dropout attacks. These findings position SecHeto-FL as a practical and scalable FL framework for next-generation intelligent networks that remains secure in the face of communication security risks.