The rapid evolution of beyond fifth-generation (B5G) and sixth-generation (6G) networks is reshaping mobile edge computing (MEC) to support large-scale, heterogeneous Internet of Things (IoT) deployments and complex cyber-physical systems (CPS). Conventional data-driven intelligence in MEC traditionally relies on centralized learning paradigms that often fail to meet the privacy, latency, scalability, and adaptability requirements in distributed and resource-constrained environments. To address these shortcomings, the objective of our work in this paper is to investigate the paradigm shift toward distributed learning and demonstrate how its co-design with emerging communication and system-level technologies can enable scalable and trustworthy intelligence for next-generation IoT and CPS. Building on this objective, we conduct a systematic survey of recent studies and analyze twelve key enabling technologies, including concept drift adaptation, transformers, TinyML, blockchain, integrated sensing and communication (ISAC), digital twins, explainable AI, federated learning and unlearning, adversarial ML, meta-learning, and multi-armed bandits. The surveyed literature is organized using a unified taxonomy and an integrated conceptual pipeline, which clarifies how these enablers interact across sensing, communication, computation, trust, and adaptation layers of IoT and CPSs. The main outcomes of this study include: (i) a comprehensive taxonomy characterizing enabling technologies for distributed edge intelligence, (ii) a comparative synthesis of representative works highlighting common architectural patterns and evaluation practices, and (iii) the identification of research gaps, critical trade-offs, and open challenges, particularly related to model robustness, energy efficiency, data heterogeneity, and secure real-time inference. Overall, this survey establishes a structured foundation and forward-looking roadmap for designing scalable, privacy-preserving, and intelligent distributed learning systems in future B5G- and 6G-enabled IoT and CPS environments.

The evolution toward 6G wireless networks demands ultra-dense cell-free massive MIMO (Multiple-Input Multiple-Output) systems that can deliver unprecedented connectivity while preserving data privacy and enabling rapid adaptation to dynamic conditions. Current resource allocation approaches rely on centralized deep reinforcement learning frameworks that create scalability bottlenecks, require extensive training periods, and violate emerging privacy regulations through global data aggregation. This paper introduces a Privacy-Preserving Federated Meta-Learning (PP-FML) framework that addresses these fundamental limitations through distributed intelligence and instant adaptation mechanisms. The proposed approach enables each access point to learn optimal resource allocation policies locally while collaboratively improving system-wide performance through cryptographically secure gradient sharing. The meta-learning component provides few-shot adaptation capabilities, allowing networks to respond to new conditions within minutes rather than hours. Comprehensive performance evaluation demonstrates that PP-FML achieves 42.3% sum rate improvement, 28.8% better energy efficiency (15.2 bits/Hz/J), sub-minute adaptation latency (51 seconds), and strong privacy guarantees epsilon 1.0 differential privacy compared to centralized approaches while maintaining complete data privacy and enabling rapid adaptation to changing network conditions. The framework scales linearly to ultra-dense deployments exceeding 300 access points per square kilometer with constant per-node computational complexity, making it suitable for practical 6G network deployment with heterogeneous device populations including mobile users and diverse applications.

The use of healthcare data for collaborative machine learning training amplifies the demand for strong privacy and security protection. Federated learning (FL) addresses this by sharing model gradients instead of raw patient data. However, FL faces critical challenges, including vulnerability to adversarial attacks (e.g., membership inference, model poisoning), reliance on centralized aggregation (introducing single points of failure), and privacy leakage risks from gradient exchanges. To overcome these limitations, we propose C2SecFL, a cross-clustered secure FL framework that combines adaptive, model-aware clustering with an inner-product functional-encryption (IPFE) scheme enabling encrypted aggregation without a key distribution center (KDC). A lightweight permissioned blockchain provides tamper-evident coordination and auditability across clusters without introducing a centralized trust anchor. Experiments on diverse models show substantially lower cryptographic overhead compared to homomorphic-encryption baselines. Encryption is approximately 3.6× faster, and aggregation is up to 2.2× faster, while predictive utility is preserved. Under adversarial settings with compromised participants, C2SecFL reduces the average attack success rate by about 80% relative to a FedAvg baseline. It mitigates membership-inference risk by encrypting updates end-to-end and revealing only aggregate inner products. Our extensive experiments on a heart attack dataset demonstrate that C2SecFL provides a practical, auditable, and resilient approach to secure healthcare FL.

The advancement of 5G and NextG networks through Open Radio Access Network (O-RAN) architecture marks a transformative shift towards more virtualized, modular, and disaggregated configurations. A critical component within this O-RAN architecture is the RAN Intelligent Controller (RIC), which facilitates the management and control of the RAN through sophisticated machine learning-driven software microservices known as xApps. These xApps rely on accessing a diverse range of sensitive data from RAN and User Equipment (UE), stored in the near Real-Time RIC (Near-RT RIC) database. The inherent nature of this shared, multi-vendor, and open environment significantly raises the risk of unauthorized sensitive RAN/UE data exposure. In response to these privacy concerns, this paper proposes a privacy-preserving zero-trust RIC (dubbed as, ZT-RIC) framework that preserves RAN/UE data privacy within the RIC platform (i.e., shared RIC database, xA$p$p, and E2 interface). The underlying idea is to employ a computationally efficient cryptographic technique called Inner Product Functional Encryption (IPFE) to encrypt the RAN/UE data at the base station, thus, preventing data leaks over the E2 interface and shared RIC database. Furthermore, ZT-RIC customizes the xAp$p$'s inference model by leveraging the inner product operations on encrypted data supported by IPFE to enable xAp$p$to make accurate inferences without data exposure. For evaluation purposes, we leverage a state-of-the-art InterClass xApp, which utilizes RAN key performance metrics (KPMs) to identify jamming signals within the wireless network. Prototyping on an LTE/5G O-RAN testbed demonstrates that ZT-RIC not only ensures data privacy/confidentiality but also guarantees a desired model accuracy of 97.9% in detecting jamming signals as well as meeting stringent sub-second timing requirement with a round-trip time (RTT) of 0.527 seconds.

The integration of the Internet of Things (IoT) in healthcare has enabled intelligent and early heart attack detection (HAD) through artificial intelligence; however, the existing approaches suffer from high computational complexity and suboptimal performance, leading to inaccurate predictions, which can jeopardize patient well-being. Moreover, they fail to provide secure bidirectional communication between patients and medical centers while safeguarding patient privacy. Therefore, this paper addresses these limitations by proposing a novel secure and efficient HAD approach in healthcare. First, we propose a customized consortium blockchain network that leverages group signatures to ensure patient anonymity, data unlinkability, and secure two-way communication, thereby preserving patient privacy. Then, a lightweight, robust HAD model is devised via knowledge distillation by leveraging a novel proposed hybrid deep learning architecture that enables accurate early detection of heart attacks, supporting timely clinical intervention. Experimental results on a real dataset, i.e., the Cleveland dataset, demonstrate the scalability of the proposed approach that can process up to 500,000 patients in under 2.5 minutes, while preserving patient privacy. Moreover, it offers 99.22% accuracy, an F1-score of 99.23%, outperforming state-of-the-art techniques, with an inference time of 90 ms and a model memory footprint of only 0.28 MB.

Healthcare is revolutionized by the integration of the Internet of Medical Things (IoMT) and artificial intelligence (AI), enabling real-time patient monitoring, advanced predictive analytics, and personalized treatment plans. However, the existing AI healthcare models are typically trained offline on static datasets, limiting their adaptability to the dynamic nature of health data. This may result in compromising models' accuracy and healthcare decision-making, rendering them obsolete. Moreover, attackers may exploit concept drift by injecting frequent data shifts, which can exhaust healthcare institutions' resources. To address this research gap, we propose a novel adaptive, secure, and efficient concept drift detection framework for healthcare. First, a robust deep learning (DL) model is devised to leverage its high-confidence probability to detect data drift efficiently without relying on labeled data. Then, we propose a customized consortium blockchain network that leverages group signatures to ensure anonymity and unlinkability of patients' health data. It also utilizes a dualledger structure, facilitating a unified drift detection model and enabling authenticated, drift-specific data sharing among medical centers. This design protects against data tampering and falsely claiming drift incidents. Our experiments, conducted on a real health monitoring dataset, show that our concept drift detection approach achieves comparable drift detection performance to the existing methods while reducing the computational time by 52.35%, and achieving an accuracy of 98.43 with our offline model and a 95% accuracy with the online adaptive model.

With the rise of Software-Defined Networking (SDN) for managing traffic and ensuring seamless operations across interconnected devices, challenges arise when SDN controllers share infrastructure with deep learning (DL) workloads. Resource contention between DL training and SDN operations, especially in latency-sensitive IoT environments, can degrade SDN's responsiveness and compromise network performance. Federated Learning (FL) helps address some of these concerns by decentralizing DL training to edge devices, thus reducing data transmission costs and enhancing privacy. Yet, the computational demands of DL training can still interfere with SDN's performance, especially under the continuous data streams characteristic of IoT systems. To mitigate this issue, we propose REDUS (Resampling for Efficient Data Utilization in Smart-Networks), a resampling technique that optimizes DL training by prioritizing misclassified samples and excluding redundant data, inspired by AdaBoost. REDUS reduces the number of training samples per epoch, thereby conserving computational resources, reducing energy consumption, and accelerating convergence without significantly impacting accuracy. Applied within an FL setup, REDUS enhances the efficiency of model training on resource-limited edge devices while maintaining network performance. In this paper, REDUS is evaluated on the CICIoT2023 dataset for IoT attack detection, showing a training time reduction of up to 72.6% with a minimal accuracy loss of only 1.62%, offering a scalable and practical solution for intelligent networks.

Ensuring data privacy in blockchain systems remains challenging due to the heavy computational and communication costs of traditional cryptographic mechanisms. Existing solutions often suffer from limited scalability, high resource consumption, and inefficient tamper-proof key management. To address these challenges, we propose Zero Trust Privacy (ZTP), a lightweight framework for scalable on-chain privacy and secure distributed key management. ZTP introduces a hybrid quorum protocol using dynamic scale-free graph adjustments and a parallel data and key management mechanism based on the Geometric Fragmentation Technique (GFT), achieving efficient, tamper-resistant shard handling. To further enhance scalability, we incorporate a lightweight consensus protocol with parallel transaction processing, isolating transactions and key access from untrusted blockchain nodes. We implement and evaluate ZTP on a distributed blockchain prototype, demonstrating outstanding performance, achieving up to 49% fault tolerance, and delivering speedups of at least 55 × compared to state-of-the-art blockchain protocols. Our results highlight ZTP’s potential for resource-constrained and large-scale blockchain deployments.

The healthcare sector consistently handles vast amounts of sensitive data, which must always be kept secure and private. Unlike classical machine learning (ML) models that require centralizing all data, federated learning (FL) addresses privacy concerns by enabling collaborative learning without sharing raw data. However, FL models are subject to limitations when managing diverse datasets from pervasive sources and encounter challenges such as adversarial threats; in addition, the generalizability of the resulting global models may limit the effectiveness of the analysis. This paper presents a novel decentralized and cluster-based FL framework designed to enhance healthcare data privacy and strengthen the security of FL processes. This framework addresses vulnerabilities by decentralizing the FL models. It includes a weighted voting mechanism that aims to improve analytics accuracy by aggregating decisions from multiple clusters. Additionally, the proposed approach reduces the risk of adversarial attacks by employing both cluster-based strategies and feature-squeezing (FS). Experimental results show that our approach surpasses classical FL methods in accuracy and security at various stages of the FL process.

Network slicing enables multiple virtual networks on shared 6G infrastructure, but dynamic resource allocation across Enhanced Mobile Broadband (eMBB), Ultra-Reliable Low Latency Communications (URLLC), and Massive Machine Type Communications (mMTC) services remains challenging. We present a hybrid Convolutional Neural Network-Long Short-Term Memory Architecture (CNN-LSTM) framework with service-specific utility functions that optimize resources while ensuring Quality of Service (QoS) guarantees under dynamic conditions. Our approach integrates spatial pattern recognition with temporal prediction, incorporating constraint measurement and lightweight optimization. Experimental results on a testbed with 100 base stations and 10,000 users demonstrate superior performance over state-of-the-art methods. The framework achieves significant improvements in resource utilization, QoS satisfaction, and energy efficiency with real-time inference capability. Convergence analysis validates system stability, confirming practical deployment feasibility for latency-critical 6G applications.

The escalating integration of deep neural networks (DNNs) in autonomous vehicles underscores the urgency of fortifying them against adversarial attacks. This paper presents a novel approach to enhance the robustness of convolutional neural networks (CNNs) in self-driving cars through a combination of adversarial mitigation techniques: they are randomization, image padding, and, most uniquely, the addition of random Gaussian noise after convolution layers. Our specialized neural network demonstrates consistent steering control under various attack scenarios, avoiding the over-steering or under-steering issues observed in standard models. As 6 G networks emerge with their ultra-reliable low-latency communication capabilities, our research contributes to the security foundation necessary for autonomous vehicles in this coming era, where resilience against adversarial manipulation will be crucial for maintaining safety in increasingly connected transportation ecosystems. Our open-sourced model provides a benchmark for real-time attackresistant systems applicable to 6G-enabled autonomous driving technologies.

Autonomous vehicles generate massive sensor data daily but operate as isolated intelligence units due to privacy constraints and network limitations. Current centralized machine learning approaches face critical barriers including compliance issues, high bandwidth costs, and latency constraints preventing real-time safety decisions. While Federated Learning (FL) enables collaborative training without raw data sharing and 6G networks promise ultra-low latency, a fundamental mismatch exists between FL's dynamic computational demands and 6G's static resource allocation mechanisms. This paper presents Predictive FL-Aware Network Slicing (PFANS), a novel framework that integrates real-time convergence modeling with proactive 6 G slice reconfiguration for autonomous vehicle networks. PFANS predicts FL computational demands multiple training rounds in advance and automatically reconfigures network slices before bottlenecks occur. Experimental results demonstrate superior resource utilization efficiency, significantly faster convergence compared to baseline approaches, and excellent handover success rates with minimal context migration times. The framework achieves state-of-theart prediction accuracy while introducing negligible network overhead, establishing effective adaptive resource management for next-generation vehicular networks.

Integrating the Internet of Medical Things (IoMT) and artificial intelligence (AI) is revolutionizing healthcare by enabling real-time health monitoring, predictive analytics, and personalized treatment. However, existing AI healthcare models are trained offline on static datasets, making them less adaptable to evolving health data and potentially reducing their accuracy and decision-making. Furthermore, adversaries may exploit this by injecting frequent data shifts, straining healthcare resources. Privacy concerns also arise from the exposure of sensitive patient data. Therefore, we propose a novel AI-driven healthcare methodology with secure, decentralized, and privacy-enhancing adaptive intelligence. First, a deep learning (DL) model is devised to leverage its high-confidence probability to detect data drift efficiently. Next, we propose a privacy-preserving approach leveraging functional encryption to ensure patient data confidentiality during drift detection and model retraining while eliminating reliance on a trusted entity. Lastly, we propose a customized consortium blockchain with group signatures to protect patient anonymity and data tampering and unlinkability while preventing falsely claiming drift incidents. Moreover, to ensure decentralization, it removes the need for a trusted authority in cryptographic key generation. Our experiments, on a real testbed and healthcare datasets, show that the proposed methodology achieves real-time drift detection with performance comparable to existing methods, while reducing the computational time by 52.35%. It also maintains high accuracy, achieving up to 98.43% with the offline health monitoring model and up to 96% with the online adaptive model. Additionally, it preserves patient privacy while reducing computational and communication overhead by 94.26% and 89%, respectively, compared to the state-of-the-art.

Heart attack detection (HAD) and human activity recognition (HAR) rely on wearable sensor data to track heart health and physical activity in real-time, facilitating early detection and monitoring of health issues. However, existing solutions for HAR and HAD face challenges in effectively capturing spatial features, long-term dependencies, and diverse sensor data representations. These shortcomings impact recognition accuracy, memory efficiency, and processing speed, while also demanding substantial computational resources due to their complexity. They also lead to performance degradation, increasing the risk of inaccurate diagnoses and potentially jeopardizing patient lives. To overcome these limitations, a novel lightweight hybrid architecture for HAR and HAD is proposed, leveraging convolutional neural networks (CNNs) with gated recurrent units (GRUs) and multi-head attention (MHA). CNNs capture spatial features, GRUs extract long-term dependencies, and MHA computes attention weights across data segments to highlight the most relevant features for health diagnosis, ensuring both improved performance and practicality for real-time health monitoring. Moreover, a magnitude-based weight pruning technique is adapted to reduce the proposed architecture’s complexity, making it suitable in resource-constrained settings without sacrificing accuracy. Furthermore, our methodology integrates an optimized genetic algorithm for counterfactual explanations, recommending minimal health data changes to lower heart attack risk. Experimental results on a real testbed and datasets, including PAMAP2, WISDM, and Cleveland, demonstrate that the proposed method outperforms the state-of-the-art methods, achieving up to 3% improvement in F1-score and accuracy, while reducing inference time, number of parameters, and memory footprint by over 40%, 70%, and 60%, respectively.

Unmanned Aerial Vehicles (UAVs) play a crucial role in various applications, including detecting environmental hazards, e.g., wildfire smoke detection. However, the limited computational capabilities and battery life of UAVs present barriers to deploying complex artificial intelligence (AI) models onboard. To address this challenge, we propose a novel hybrid deep learning framework for UAVs to carry out light-weight yet efficient smoke detection. The framework combines a lightweight model for initial image assessment and a depth-wise model for selective processing of uncertain cases. Bayesian optimization is employed to determine the optimal threshold values for activating the depth-wise model, striking a balance between accuracy and computational efficiency. The proposed approach eliminates the need for cloud server connectivity, enabling onboard decision-making. Experimental results demonstrate that the hybrid framework achieves significant reductions in processing time and the number of calls to the depth-wise model while maintaining high accuracy. The framework’s adaptability and robustness make it suitable for real-time smoke detection applications in resource-constrained environments.

The convergence of Intelligent Reflecting Surfaces (IRS) and Terahertz (THz) communications represents a transformative advancement for sixth-generation (6G) wireless networks, yet presents unprecedented challenges in system optimization. This paper addresses the critical challenge of joint optimization between IRS phase shifts and THz resource allocation in dynamic Internet of Things (IoT) environments, focusing on real-time adaptation to rapidly changing channel conditions. We propose a novel Adaptive Online Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework that leverages dynamic experience weighting to automatically adjust learning based on detected environmental changes. Our approach incorporates a multi-resolution buffer structure that balances recent observations with historical patterns, enabling both rapid adaptation and long-term optimization while considering the unique characteristics of THz-band propagation and IRS reflection patterns. The framework employs explicit coordination protocols between IRS controllers and resource managers, significantly improving convergence in non-stationary environments. Comprehensive simulations using realistic THz channel models and practical IRS configurations demonstrate that our proposed framework achieves a 45% improvement in system throughput, a 38% reduction in end-to-end latency, and a 30% enhancement in energy efficiency compared to conventional optimization approaches. More significantly, our solution demonstrates unprecedented adaptation capabilities, recovering 90% of optimal performance within 5 ms after abrupt environmental changes a critical requirement for future 6G networks. The framework maintains robust performance under diverse conditions, including high user mobility scenarios and adverse atmospheric conditions, while exhibiting linear computational scaling with increasing IRS elements (tested up to 512 elements). These results establish the viability of Adaptive Online MADDPG-based joint IRS-THz optimization for practical 6G deployments, particularly in dynamic IoT environments where traditional communication approaches face significant limitations.

Cloud-based federated learning (FL) services have received increasing attention due to their ability to enable collaborative global model training without the need to collect local data from participants. To generate a global model, local models are trained on participants' local data and only model parameters are sent to an aggregator server. Nonetheless, revealing model parameters can still reveal training data via launching attacks, e.g., inference and membership. Hence, to protect model parameters, a secure global model aggregation scheme is needed to protect these parameters from unauthorized access. Existing solutions to this issue, which are based on homomorphic encryption and secure multi-party computation, tend to have large overheads and slow down training times. Functional encryption (FE) has been proposed as a solution for resolving privacy-preservation issues in FL, but current solutions suffer from high overhead and lack of security such as leaking master private key. To address these issues, this paper proposes a privacy-protecting, efficient, and decentralized FL framework, called FedSafe, based on FE without the need for a trusted key distribution center (KDC). The proposed scheme allows the participants to communicate with an aggregator to construct a global model without disclosing or learning their local models' parameters or the training data, thereby safeguarding their privacy. Through rigorous testing with real-world data, it is demonstrated that FedSafe outperforms the state-of-the-art privacy-protecting FL schemes in terms of security, scalability, and communication and computation overhead. Unlike existing approaches, this is accomplished without depending on any trusted KDC.

Artificial Intelligence (AI) applications on Internet of Things (IoT) networks often involve relaying generated data to a server for deep learning training, which poses security risks to users' data. Federated Learning (FL) offers a distributed model training paradigm in which local data are kept at the edge and locally trained models are exchanged and aggregated by a server over several rounds to produce a global model. While successful, standard FL algorithms do not support heterogeneous local model design, an essential requirement, especially for resource-limited edge devices. Recently, Knowledge Distillation-based FL algorithms have provided model-agnostic FL to enable clients to independently design their local model and share soft labels instead of model parameters. KD-based FL algorithms are computationally expensive due to additional distillation training. We propose Federated Learning with Selective Knowledge Distillation (FedSKD) to address the limitations of system heterogeneity; and computation and communication demands. We evaluate different aspects of the proposed algorithm relative to baseline FL algorithms. Results show that FedSKD incurs significantly less per-round computation time and communication overhead relative to the considered model-based and KD-based FL algorithms.

Load forecasting (LF) in smart grids is beneficial not only in mitigating equipment failures and power outages but also in facilitating effective power dispatching and infrastructure planning. To predict future loads accurately, the consumers' fine-grained energy consumption readings are fed into machine-learning (ML) models. However, revealing these readings enables adversaries to deduce confidential information about consumers, including details about their lifestyle, and hence their privacy is violated. To address this privacy issue, the existing works only focus on using federated learning (FL)-based approaches to train and obtain an accurate global LF model. Nevertheless, addressing the privacy violation problem during the LF process (in the deployment phase) after obtaining the global model for AMI networks has not been well investigated yet. Therefore, this paper proposes a novel, efficient, and decentralized approach that enhances the precision of LF while safeguarding the privacy of consumers. The proposed scheme incorporates inner product functional encryption (IPFE) to allow smart meters (SMs) to encrypt their readings with no need for a trusted key distribution center (KDC) while allowing LF without divulging or acquiring knowledge of the consumers' readings to protect their privacy. In addition, a hybrid deep learning approach is developed to construct an LF model that can yield precise forecasts. To show the feasibility of the proposed scheme, the performance of our scheme was assessed on a real energy consumption readings dataset, and the results demonstrate proficiency in LF while providing robustness and privacy preservation with reasonable communication efficiency.

In federated learning (FL) based electricity theft detection, detection nodes (DNs) locally train deep learning models on consumers' data and share only the local model parameters with an aggregation server (AS) to generate a global model shared by all nodes for better detection accuracy. However, several privacy concerns should be addressed including membership and inference attacks. To mitigate these attacks, several privacy-preserving aggregation schemes have been introduced. Nevertheless, existing FL detectors often overlook the threat of poisoning attacks, in which certain DNs hold maliciously labeled, i.e., poisoned, data during the training. This manipulated data can subsequently be exploited to introduce backdoors into the global model after its deployment. This paper introduces a novel approach that enhances privacy and resilience against poisoning attacks in FL-based electricity theft detection within smart grids. Our approach enables encrypting local parameters before sending them to the AS, thus safeguarding consumers' privacy. Additionally, it utilizes a cosine similarity test over encrypted data to detect and mitigate poisoning attacks by filtering out malicious local gradients from being considered in the global model computation. Through extensive evaluations, we demonstrate the effectiveness of our FL-based detector in substantially reducing the poisoning attack success rate even when 50% of DNs train their local models with malicious targeted power consumption data, all while preserving consumers' privacy.

Human activity recognition (HAR) leverages data from wearable devices and smartphones to detect actions, improving quality of life in areas like elderly care, health monitoring, and sports training. Current deep learning architectures struggle with extracting spatial features, long-term dependencies, and diverse sensor data representations, impacting recognition performance and posing challenges for resource-constrained IoT devices due to their complexity and parameter count. We propose a novel hybrid HAR architecture, integrating convolutional neural networks (CNN) and gated recurrent units (GRU) with a multi-head attention (MHA) mechanism. This architecture captures spatial features via CNN, extracts long-range dependencies with GRU, and uses MHA to compute attention weights for different data segments. The combined spatial and attention features are fed into a classification module for activity recognition. On the PAMAP2 dataset, our CNN-GRU-MHA model outperforms existing methods, achieving an F1-score of 98.4 %, with an inference time of 0.078 seconds and a memory footprint of 790.02 KB, reducing resource usage by 74.34 % and 62.81 %, respectively.

In the realm of smart grids, smart meters can be hacked to report false data to lower the consumers’ electricity bills. While machine learning (ML) techniques have shown promise in detecting false data, they are also prone to adversarial attacks, such as evasion attacks. This article investigates the impact of gradient-ensemble-based evasion attacks on the smart grid ML-based false data detectors, focusing on the white-box threat model where attackers possess detailed knowledge of the defense mechanism. First, we examines the vulnerability of three detectors (consumer-based, cluster-based, and global) to gradient-based evasion attacks. The evaluation results show an inverse relationship between robustness of the detectors and regularization (i.e., generalization), where higher data set variability usually causes higher regularization. Notably, minimal regularization level is observed when electricity consumption patterns are close. Our findings also indicate that the consumer-based detector exhibits higher accuracy and robustness but remains susceptible to zero day attacks and demands substantial computational resources for training an ML model for each consumer. In contrast, the cluster-based detector improves accuracy and exhibits satisfactory robustness compared to the global detector. Subsequently, we proposes two parallel-ensemble approaches (stacking and voting) for the cluster-based false data detectors trained on the adversarial samples. The evaluation results demonstrate that integrating clustering, adversarial training, and ensemble methods, the proposed detector enhances robustness against gradient-ensemble-based evasion attacks while significantly boosting accuracy. This stands in contrast to benchmark defenses, which often face a tradeoff between accuracy and robustness, sacrificing accuracy to bolster resilience against evasion attacks.

Load forecasting (LF) is a crucial process of predicting future energy load and demand in smart grids, allowing for mitigating equipment failures and power outages, besides facilitating effective power dispatching and infrastructure planning. Methods used in LF range from traditional statistical and mathematical models to modern machine learning (ML) algorithms, and it has been proven that the latter has a better ability to predict future loads. These techniques leverage the consumers’ energy consumption readings for use in the LF process; however, revealing these readings exposes sensitive consumer lifestyle information, hence violating their privacy. The majority of the existing works focus on obtaining precise LF and addressing consumers’ privacy violations during the LF process in the deployment phase has not been well-investigated yet. Moreover, the existing methods that can be employed to preserve privacy introduce high overhead and rely on a trusted third party, which undermines the trust assumption, making them less robust. Therefore, this article proposes a novel and efficient privacy-preserving LF scheme, called privacy-preserving LF and monitoring and billing (PLFMB), that utilizes inner product functional encryption (IPFE) and eliminates the need for a trusted key distribution center. PLFMB allows smart meters to encrypt their consumption readings while enabling the system operator (SO) to 1) evaluate a hybrid deep learning-based LF model developed to predict future loads accurately; 2) monitor the grid load; and 3) compute consumers’ bills following dynamic pricing, without revealing or learning consumers’ readings to protect their privacy. Our proposed scheme has been evaluated on a real energy consumption data set, demonstrating its feasibility, proficiency in LF, and robustness in preserving consumer’s privacy while maintaining reasonable overhead.

Federated learning is a promising approach for training models on distributed data, driven by increasing demand in various industries. However, federated learning framework faces several key challenges, including communication bottlenecks and client data heterogeneity. Personalized asynchronous federated learning addresses these challenges by customizing the model for individual users based on their local data while trading model updates asynchronously. In this paper, we propose the Personal-ized Moreau Envelopes-based Asynchronous Federated Learning (APFedMe). Our approach combines the strengths of Moreau En-velopes to handle optimization problems and asynchronous weight updates to improve communication efficiency while mitigating heterogeneity data challenges through a personalized learning environment. We evaluate our approach on several datasets and compare it with the baseline PFedMe method. Our experiments demonstrate that the proposed APFedMe outperforms other meth-ods in terms of convergence speed and communication efficiency. Overall, our work contributes to developing more effective and efficient federated learning methods that can be applied in various real-world scenarios.

Satellite technologies have advanced drastically in recent years, leading to a heated interest in launching small satellites into low Earth orbit (LEOs) to collect massive data such as satellite imagery. Downloading these data to a ground station (GS) to perform centralized learning to build an AI model is not practical due to the limited and expensive bandwidth. Federated learning (FL) offers a potential solution but will incur a very large convergence delay due to the highly sporadic and irregular connectivity between LEO satellites and GS. In addition, there are significant security and privacy risks where eavesdroppers or curious servers/satellites may infer raw data from satellites' model parameters transmitted over insecure communication channels. To address these issues, this paper proposes FedSecure, a secure FL approach designed for LEO constellations, which consists of two novel components: (1) decentralized key generation that protects satellite data privacy using a functional encryption scheme, and (2) on-orbit model forwarding and aggregation that generates a partial global model per orbit to minimize the idle waiting time for invisible satellites to enter the visible zone of the GS. Our analysis and results show that FedSecure preserves the privacy of each satellite's data against eavesdroppers, a curious server, or curious satellites. It is lightweight with significantly lower communication and computation overheads than other privacy-preserving FL aggregation approaches. It also reduces convergence delay drastically from days to only a few hours, yet achieving high accuracy of up to 85.35% using realistic satellite images.

Federated Learning (FL) has emerged as a powerful approach to facilitate the construction of centralized models without compromising the data privacy of multiple participants. However, conventional FL methodologies do not address system heterogeneity where each participant needs to independently design its own model, a prevalent requirement in Internet of Things (IoT) applications due to the heterogeneous nature of tasks and data. Knowledge Distillation-based FL algorithms tackle this limitation by exchanging soft labels instead of model weights, thus giving each client the ability to independently design its local model architecture. While FL is inherently private, studies have indicated that exploiting gradients for a few iterations can reveal sensitive training data. To protect against privacy attacks, FL algorithms employ Differential Privacy (DP) to guarantee privacy protection, which can be applied using Local Differential Privacy (LDP). In this paper, we elaborate on preserving clients' training data privacy in KD (Knowledge Distillation)-based FL using DP, providing both privacy and flexibility. We provide theoretical analysis to extend the privacy guarantee to exchanged updates. Experimental analysis is performed utilizing Human Activity Recognition (HAR) datasets with different modalities. The results obtained demonstrate the capacity of KD-based FL to maintain a robust utility-privacy balance. Furthermore, for the same DP protection level, the utility of models trained on images was severely reduced across all FL algorithms. This suggests that the modality and complexity of a dataset are important factors for shaping the utility-privacy tradeoff of DP.

Energy forecasting is important because it enables infrastructure planning and power dispatching while reducing power outages and equipment failures. It is well-known that federated learning (FL) can be used to build a global energy predictor for smart grids without revealing the customers’ raw data to preserve privacy. However, it still reveals local models’ parameters during the training process, which may still leak customers’ data privacy. In addition, for the global model to converge, it requires multiple training rounds, which must be done in a communication-efficient way. Moreover, most existing works only focus on load forecasting while neglecting energy forecasting in net-metering systems. To address these limitations, in this article, we propose a privacy-preserving and communication-efficient FL-based energy predictor for net-metering systems. Based on a data set for real power consumption/generation readings, we first propose a multidata-source hybrid deep learning (DL)-based predictor to accurately predict future readings. Then, we repurpose an efficient inner-product functional encryption (IPFE) scheme for implementing secure data aggregation to preserve the customers’ privacy by encrypting their models’ parameters during the FL training. To address communication efficiency, we use a change and transmit (CAT) approach to update local model’s parameters, where only the parameters with sufficient changes are updated. Our extensive studies demonstrate that our approach accurately predicts future readings while providing privacy protection and high communication efficiency.

Energy theft causes economic losses and power out-ages and disrupts energy generation and distribution of smart grids. A significant challenge is how to effectively use customers' power consumption data for energy theft detection while pre-serving security and privacy. One solution is to use federated learning (FL) to compute a global model to detect energy theft cyberattacks where detection stations train local models on their customers' power consumption data and send only the parameters of the models to an aggregator server. Nevertheless, revealing the model's parameters may still leak customers' private data by launching attacks such as membership and inference. Therefore, a secure aggregation scheme is needed to protect the models' param-eters. Furthermore, the existing privacy-preserving aggregation schemes suffer from high overhead and low model accuracy. This paper addresses these limitations by proposing a novel privacy- preserving, efficient, decentralized, aggregation scheme based on a functional encryption cryptosystem for energy theft detection in smart grids without requiring a key distribution center. Our scheme enables the detection stations to send encrypted training parameters to an aggregator, which calculates the aggregated parameters and returns the updated model parameters to the detection stations without being able to learn the parameters of the local models or the training data of the customers to preserve their privacy. Moreover, the results of our extensive experiments show that our FL-based detector can detect energy thefts accurately with low overhead because of our lightweight privacy-preserving aggregation scheme.

In the smart grid, malicious customers may compromise their smart meters (SMs) to report false readings to achieve financial gains illegally. This causes hefty financial losses to the utility and may degrade the grid performance because the reported readings are used for energy management. This article is the first work that investigates this problem in the net-metering system, in which one SM is used to report the difference between the power consumed and the power generated. First, we prepare a benign data set for the net-metering system by processing a real power consumption and generation data set. Then, we propose a new set of attacks tailored for the net-metering system to create a malicious data set. After that, we analyzed the data and found time correlations between the net meter readings and correlations between the readings and relevant data obtained from trustworthy sources, such as solar irradiance and temperature. Based on the data analysis, we propose a general multidata-source deep hybrid learning-based detector to identify the false-reading attacks. Our detector is trained on net meter readings of all customers besides data from trustworthy sources to enhance the detector performance by learning the correlations between them. The rationale here is that although an attacker can report false readings, he cannot manipulate the solar irradiance and temperature values because they are beyond his control. Extensive experiments have been conducted, and the results indicate that our detector can identify the false-reading attacks with a high detection rate of 98.59% and a low false alarm of 2.92%.

The periodic transmission of the customers’ power consumption readings in the advanced metering infrastructure (AMI) is essential for energy management and billing. To collect the readings efficiently, the change and transmit approach is adopted in AMI (CAT AMI) so that the readings are reported only when there is enough change in the consumption. However, CAT AMI suffers from malicious customers who launch electricity-theft cyberattacks by manipulating their readings to illegally reduce their bills. These attacks can cause hefty financial losses and degrade the grid performance because the readings are used for grid management. In this article, the electricity-theft problem in CAT AMI networks is investigated. We first process a real power consumption readings data set to create a benign data set and propose a new set of cyberattacks to create malicious samples. We then develop a deep-learning-based electricity-theft detection solution to identify malicious customers for the CAT AMI network. The proposed detector uses both the customers’ transmission pattern and CAT readings to learn the correlation between them in order to enhance the detector’s ability in identifying electricity thefts. We conduct extensive experiments to evaluate the performance of our electricity-theft detector, and the results indicate that our detector can accurately detect malicious customers and achieve higher detection rate and lower false alarm than the detectors that are trained only on the CAT readings.

There are different metering systems adopted in the advanced metering infrastructure (AMI) of the smart grid. Among these systems, the net-metering is a promising system that motivates customers to install renewable resources at their premises to generate electricity and sell it to the utility. In this system, the customer’s home is equipped with one smart meter to report the net readings representing the difference between the power consumed from the power grid and the power injected into the grid. However, malicious customers may compromise their meters to report false readings to the utility to illegally achieve financial gains. This not only causes huge losses to the utility, but also deteriorates the grid performance. To the best of our knowledge, this problem has not been investigated. Therefore, in this paper, we investigate the detection of false-reading attacks in the net-metering system for the first time. Specifically, we propose four sophisticated attacks customized for the net-metering system and use them to create a dataset containing both benign and malicious samples. We have analyzed the dataset and detected time correlations between the readings within the benign samples. Based on the data analysis, we propose a general deep-learning-based detector with hybrid architecture involving convolutional neural network (CNN) and gated recurrent unit neural network (GRU). We have evaluated our detector, and the results demonstrate that the detector can detect the false-reading attacks with high precision and recall, and low false alarm.

There are two power consumption readings collection approaches adopted in the advanced metering infrastructure (AMI) of the smart grid; periodic transmission (PT) and change and transmit (CAT) AMI systems. Among these approaches, CAT is a promising approach that collects these readings efficiently by sending the readings only when there is enough change in consumption to reduce the number of transmitted readings. However, CAT AMI system suffers from electricity theft cyber-attacks that can be launched by malicious customers who may compromise their meters and manipulate their power consumption readings to illegally reduce their bills. These attacks do not only cause hefty financial losses but may also degrade the grid performance because the readings are used for grid management. Therefore, this paper is the first work that investigates this problem for CAT AMI system, in which the power consumption readings are not sent periodically to the system operator. We first prepare a benign dataset for the CAT AMI by processing a real power consumption readings dataset. Next, we propose a new set of attacks tailored for the CAT AMI to create a malicious dataset. Then, we propose a general and hybrid deep-learning electricity theft detector to identify malicious customers. The proposed detector is trained on both benign and malicious data from all customers using the reported CAT readings. Extensive test studies are carried out to investigate the detector’s performance using publicly available real data of power consumption from 114 customers. Simulation results demonstrate our models can detect malicious customers with high detection rate and low false alarm.

Reporting fine-grained power consumption readings periodically in advanced metering infrastructure (AMI) results in transmitting a massive amount of data by each smart meter (SM). To collect these readings efficiently, change and transmit (CAT) approach can be used. In CAT, the SM sends a consumption reading only when there is enough change in the consumption, which reduces the number of transmitted readings. However, using the CAT approach may trigger attackers to launch a presence-privacy attack (PPA) to infer sensitive information such as the absence of the house occupants by analyzing their SM’s transmission pattern. Therefore, in this paper, we propose a scheme, called “STID”, for collecting the power consumption readings efficiently in AMI networks while preserving the consumers’ privacy by transmitting spoofing transmissions based on an interactive deep-learning defense model. First, we create a dataset that contains the CAT transmission patterns using real power consumption readings and a clustering technique. Next, we train a deep-learning-based attacker model to launch PPA, and the results indicate that the success rate of the attacker is about 90%. Finally, to mitigate the PPA, we train a defense model using deep-learning to transmit spoofing transmissions. The evaluations of our envisioned STID scheme demonstrate a significant reduction in the attacker’s success rate while achieving high efficiency in terms of the number of readings that should be transmitted. Our measurements indicate that our proposed STID can reduce the attacker’s success rate to 6.12% and increase efficiency by about 38% compared to transmitting readings periodically.

In advanced metering infrastructure, smart meters (SMs) send fine-grained power consumption readings periodically to the utility for load monitoring and energy management. Change and transmit (CAT) is an efficient approach to collect these readings, where the readings are not transmitted when there is no enough change in consumption. However, this approach causes a privacy problem, that is, by analyzing the transmission pattern of an SM, sensitive information on the house dwellers can be inferred. For instance, since the transmission pattern is distinguishable when dwellers are on travel, attackers may analyze the pattern to launch a presence-privacy attack (PPA) to infer whether the dwellers are absent from home. In this article, we propose a scheme, called “STDL,” for efficient collection of power consumption readings in advanced metering infrastructure (AMI) networks while preserving the consumers’ privacy by sending spoofing transmissions using a deep-learning approach. We first use a clustering technique and real power consumption readings to create a data set for transmission patterns using the CAT approach. Then, we train a deep-learning-based attacker model, and our evaluations indicate that the attacker’s success rate is about 91%. Finally, we train a deep-learning-based defense model to send spoofing transmissions efficiently to thwart the PPA. Extensive evaluations are conducted, and the results indicate that our scheme can reduce the attacker’s success rate to 3.15%, while still achieving high efficiency in terms of the number of readings that should be transmitted. Our measurements indicate that the proposed scheme can increase efficiency by about 41% compared to continuously transmitting readings.

In advanced metering infrastructure (AMI), smart meters (SMs) are installed at the consumer side to send fine-grained power consumption readings periodically to the system operator (SO) for load monitoring, energy management, and billing. However, fraudulent consumers launch electricity theft cyber attacks by reporting false readings to reduce their bills illegally. These attacks do not only cause financial losses but may also degrade the grid performance because the readings are used for grid management. To identify these attackers, the existing schemes employ machine-learning models using the consumers' fine-grained readings, which violates the consumers' privacy by revealing their lifestyle. In this article, we propose an efficient scheme that enables the SO to detect electricity theft, compute bills, and monitor load while preserving the consumers' privacy. The idea is that SMs encrypt their readings using functional encryption (FE), and the SO uses the ciphertexts to: 1) compute the bills following the dynamic pricing approach; 2) monitor the grid load; and 3) evaluate a machine-learning model to detect fraudulent consumers, without being able to learn the individual readings to preserve consumers' privacy. We adapted an FE scheme so that the encrypted readings are aggregated for billing and load monitoring and only the aggregated value is revealed to the SO. Also, we exploited the inner-product operations on encrypted readings to evaluate a machine-learning model to detect fraudulent consumers. The real data set is used to evaluate our scheme, and our evaluations indicate that our scheme is secure and can detect fraudulent consumers accurately with low communication and computation overhead.

Preserving the customers' privacy, while collecting their power consumption for monitoring and billing, is a prime concern in an Advanced Metering Infrastructure (AMI) network of the Smart Grid (SG). In this paper, we address this concern by formally formulating the data aggregation privacy problem, and propose a uniquely crafted Privacy-Preserving Monitoring and Billing scheme using Functional Encryption, referred to as PMBFE. Our proposed PMBFE fulfills four key objectives: (i) data aggregation for billing, (ii) dynamic pricing flexibility, (iii) load monitoring with customers' privacy preservation; and (iv) analysis on how the adopted functional encryption is able to jointly perform data aggregation efficiently and guarantee privacy-preservation. Our envisioned PMBFE approach is evaluated with extensive computer-based simulations. In contrast with the widely employed homomorphic-based encryption in AMI networks, our proposed PMBFE demonstrates significant performance improvement in terms of both communication and computation overheads while guaranteeing user-data privacy. Furthermore, the conducted security analysis exhibits the robustness of our proposal against collusion and eavesdropping attacks.