As global demand for efficient network services increases, the limitations of traditional terrestrial wireless networks are becoming more apparent. Space-air-ground integrated networks (SAGIN) have emerged as a promising solution to advance next-generation network infrastructure. However, SAGIN faces significant security challenges—including the lack of a robust security architecture, trust and data reliability issues in multi-hop transmissions, and the need to enhance network performance without compromising security—rendering traditional boundary-based defenses inadequate. Therefore, we propose SECURELINK, a decentralized zero-trust architecture tailored for SAGIN, which replaces traditional perimeter defenses with a ”never trust, always verify” approach. This approach strengthens network security and flexibility through continuous verification and adherence to the principle of least privilege. SECURELINK integrates blockchain technology to establish a multi-layered security verification and data processing scheme, addressing the dynamic and decentralized features of SAGIN. Additionally, we introduce DFRIO, a zero-trust traffic offloading method based on decentralized federated reinforcement learning and blockchain, designed to enhance network performance within maintaining security. Simulation results demonstrate that our solution significantly enhances SAGIN’s defense capability without compromising network stability, outperforming the two baseline schemes by 18.3% and 42.6%, respectively.

The popularity of edge smart devices and the explosive growth of generated data have driven the development of mobile crowd sensing (MCS). Also, federated learning (FL), as a new paradigm of privacy-preserving distributed machine learning, integrates with MCS to offer a novel approach for processing large-scale edge device data. However, it also brings about many security risks. In this paper, we propose a verifiable federated learning scheme with privacy-preserving for mobile crowd sensing. In our federated learning scheme, the double-layer random mask partition method combined with homomorphic encryption is constructed to protect the local gradients and enhance system security (strong anti-collusion ability) based on the multi-cluster structure of federated learning. Also, a sampling verification mechanism is proposed to allow the mobile sensing clients to quickly and efficiently verify the correctness of their received gradient aggregation results. Further, a dropout handling mechanism is constructed to improve the robustness of mobile crowd sensing-based federated learning. Related experimental results demonstrate that our verifiable federated learning scheme is effective and efficient in mobile crowd sensing environments.

Diabetic retinopathy (DR) is a major cause of vision loss in workingage individuals, and automated detection systems can assist clinicians in early diagnosis and treatment.However, existing grading methods trained on high-quality fundus images often exhibit significant performance degradation when applied to low-quality images commonly encountered in clinical practice.To address this issue, we propose IELS-DR, a multi-task collaborative learning framework specifically designed for low-quality fundus images.The framework jointly optimizes three tasks: image enhancement, lesion segmentation, and lesion severity classification.It incorporates a dedicated image enhancement subnet tailored for low-quality images to improve clarity and facilitate downstream tasks.Additionally, we introduce a Cross-task Feature Aggregation (CFA) module, which employs a cross-task, multi-scale CNN architecture to enable deep feature sharing and interaction between lesion segmentation and DR classification tasks.The CFA module enhances multi-scale feature fusion, strengthens lesion feature representation, and promotes effective task collaboration.Experiments on the DDR dataset show that IELS-DR delivers leading performance, with an accuracy of 86.61%, precision of 83.57%, kappa of 82.36%, recall of 85.09%, and F1-score of 84.25%.Furthermore, it delivers competitive results in image enhancement and lesion segmentation tasks.Ablation studies confirm the efficacy of the proposed framework and CFA module in improving robustness for low-quality image analysis and DR grading.

Diabetic retinopathy (DR) is a complication of diabetes.Deep learning techniques play an important role in the automatic grading of DR.However, the significant morphological differences and diverse types of lesions in DR images often lead to the neglect of the dependencies between lesions, which can significantly affect the final grading results.Additionally, traditional convolutional neural networks (CNNs) often lose contextual and discriminative information during feature extraction.To address this issue, this paper proposes a Multi-Scale Contextual Cross-Attention Network (MCCNet) for DR classification.The network includes residual blocks, a Global Context Module (GCM), a Context Attention Fusion Module (CAFM), and a Multi-Scale Feature Module (MSFM).First, the residual blocks and GCM extract lesion features while considering the global contextual information of the lesions.Then, the CAFM fuses global and multi-scale contextual information through a cross-attention mechanism to reduce information loss during the convolution process.Finally, the MSFM learns the dependencies between different lesions through dilated convolutions.DR grading is performed through a classifier.Experimental results show that the proposed method achieves accuracies of 88.93% and 96.04% on the APTOS-2019 and Messidor-2 datasets, respectively, outperforming other methods.

False data injection attacks (FDIA) seriously endanger smart grid security by intentionally manipulating the measurement data critical for power system state estimation. Although existing deep learning-based detection methods can identify such attacks by modeling the spatiotemporal dependencies of the power grids, they still suffer from several limitations. In the spatial dimension, existing approaches focus only on neighboring nodes, failing to effectively capture the global topological correlations among non-adjacent nodes. In spatiotemporal feature fusion, traditional serial architectures tend to cause feature information loss and cannot achieve deep coupling of spatiotemporal features. To overcome these limitations, we propose a FDIA detection model in smart grids using Graphormer and LSTM networks. The model incorporates a dual-branch parallel feature extraction framework, whereby an optimized Graphormer architecture captures global spatial dependencies within the power grid, while LSTM networks extract temporal dynamic features. Furthermore, a novel bidirectional crossattention fusion mechanism is designed to enable interactive and deep integration of spatiotemporal features. Experimental results on standardized IEEE 14-bus and 118-bus test systems indicate that the proposed approach surpasses existing models in multiple performance indicators and exhibits strong robustness.

The technical loss of electricity caused by electricity theft from customers has been a pressing issue for utilities worldwide. With the rapid development of power equipment and artificial intelligence (AI) technology, machine learning (ML)-based electricity theft detection schemes have attracted widespread attention from power companies. However, existing ML-based power theft detection models often can only detect some simple anomalous data and cannot effectively detect fraudulent data generated by evasion attacks. To address these issues, we first use multi-user data as input to extract multi-user electricity usage features and increase the robustness of the detection model. Besides, we propose a CNN-AE-LSTM hybrid model-based electricity theft detection scheme, where the Convolutional Neural Network (CNN) model and the Autoencoder (AE) model are used to extract power data features from the temporal relationships of multi-users, the Long Short-Term Memory (LSTM) model is used to predict electricity theft by learning the relationship between the extracted power features and the temporal sequence of each electricity theft user. Finally, experiments under white-box and black-box settings show that in the face of the three evasion attacks, the detection rate (DR) of the method is 88.1%-89.3% and 92.3%-95.7%, and the false-alarm rate (FA) is 3.1%-6.6% and 2.7%-4%, which is higher than that of several other ML-based electricity theft detection methods.

Due to significant regional variations in electricity consumption, transfer learning methods for detecting electricity theft can effectively address challenges such as limited labeled data and domain shifts in emerging areas. However, these methods often encounter issues related to inadequate boundary feature characterization, underutilization of available data, and substantial computational overhead. To address these challenges, we propose a semi-supervised domain adaptation method for electricity theft detection based on minimax entropy. Specifically, domain-invariant prototype vectors align the feature distributions of labeled data across source and target regions, thereby enhancing cross-domain consistency. This alignment reduces inter-domain differences and improves detection, particularly when labeled data in target regions is scarce. Additionally, the minimax entropy strategy adjusts the prediction confidence of unlabeled data, thereby strengthening intra-class aggregation, enhancing inter-class separability, and optimizing feature distributions to capture discriminative boundary features. Experimental results demonstrate that our method significantly outperforms existing approaches in both detection performance and training efficiency.

Due to the characteristics of high mobility, low latency and short connection, task offloading in vehicular ad-hoc networks (VANETs) is facing many issues. In this paper, we propose a multi-layer task offloading scheme in fog computing-based VANETs, whose task completion delay can be optimized to an approximate minimum by a dynamical task offloading process. In our proposed scheme, a vehicle movement model is constructed to predict the position of each vehicle in a short time. Then, the link reliability evaluation metrics are constructed based on the vehicular willingness, the degree of vehicle connection and the expected count of task transmission. Furthermore, an evaluation model of processing delay and resource quantity is established to further estimate the selection of offloaded vehicles according to node degree, resource degree, computation capacity and transmission capacity. Finally, the multi-layer task offloading scheme is proposed, where we build an optimization model to minimize the task transmission and computation (completion) delays. Related experiments show our multi-layer task offloading scheme is efficient by optimizing task completion delay in fog computing-based VANETs.

With the advancement of artificial intelligence, automobiles are progressively transitioning from traditional mechanization to Connected Autonomous Vehicles (CAVs), significantly enhancing driving comfort and safety. As the standard communication protocol in CAVs, the Controller Area Network (CAN) remains vulnerable to attacks due to the lack of robust security mechanisms. While existing deep learning-based vehicle network intrusion detection systems can effectively identify known attacks, their ability to detect unknown attacks is limited due to the same data distribution in the source and target domain. To address this issue, we propose a domain adaptation and time-series imaging-based intrusion detection system (DATI-IDS) to detect known and unknown attacks, where the deep domain adaptation method is used to solve the source and target domain data distribution difference problem by optimizing the multiple kernel maximum mean discrepancy (MK-MMD) between the source domain and target domain images and the classification loss, and the time-series imaging method is used to capture temporal dependencies and improve efficiency by transforming the CAN ID sequence into a two-dimensional gramian angular summation field (GASF) image. The effectiveness of the proposed model is evaluated across nine distinct unknown attack scenarios using the Car-Hacking dataset and the survival analysis dataset. Comparative analysis with previous studies demonstrates superior performance, faster inference times, and reduced model complexity.

With the growing popularity of federated learning, the security of training models against backdoor attacks has become a key challenge. Existing defense schemes often fail to address the complexity and diversity of such attacks so as to make training models vulnerable. In this paper, we propose a comprehensive dual-layered model protection scheme for fog computing-based federated learning framework. In our scheme, we first introduce a multi-metric defense mechanism deployed on fog servers to defend against malicious backdoor attacks from edge devices. The proposed defense mechanism employs multiple detection indicators to simultaneously evaluate and detect gradient and model training attributes, so that the abnormal local gradients are identified effectively. Further, we construct a second-layered defense scheme deployed on aggregation servers to regularly monitor the participation status of fog servers, whose purpose is to detect the distribution of uploaded gradients and eliminate malicious gradients from compromised fog servers. Additionally, we design an adaptive gradient adjustment method to mitigate the influence of deleting malicious gradients on the global model training process. Experimental results show that our dual-layered model protection scheme can perform well against three type of backdoor attacks (BadNet, Blended and WaNet).

The Space-Air-Ground integrated network (SAGIN) has gained significant attention due to the explosive growth in mobile data traffic. In this network, Unmanned Aerial Vehicles (UAVs) play a critical role as air relay nodes, bridging ground and space networks. However, challenges arise from the dynamic position angles between UAVs and satellites, as well as fixed UAV altitudes, limiting air-to-space transmission capacity. Moreover, the finite UAV battery capacity carries the risk of energy interruptions during SAGIN transmissions. To address these issues, we propose an integrated model that considers UAV channel fading, energy consumption, and harvesting. This model allows us to comprehensively analyze SAGIN transmission performance. Within this framework, we calculate the UAV energy outage probability and signal-to-noise ratio (SNR) outage probability for SAGIN uplink transmission. Based on our network performance analysis, we derive an expression for the optimal UAV altitude, ensuring uninterrupted energy supply and preventing SNR outage. To assess the fairness of SAGIN transmission performance, we compare the capabilities of Ground-to-Air-to-Space and Ground-to-Space transmissions. Additionally, we provide closed-form expressions for the transmission time gap in both scenarios. Our numerical results validate the accuracy of these derived expressions and evaluate how key parameters impact the optimal UAV altitude in the SAGIN uplink.

With advancements in renewable energy technologies, microgrids have evolved with distinctive cyber-physical system (CPS) characteristics, providing a dynamic and efficient control framework. However, the susceptibility of agents to false data injection attacks (FDIAs) during information transmission poses a notable security challenge. Existing efforts focus on detecting these attacks through machine learning methods, without regard to the cyber information embedded in the CPS communications. To address this gap, we propose a cyber-physical collaborative detection method (CPCGD) based on gate recurrent unit (GRU) and deep-learning neural network (DNN) to detect FDIAs, where the GRU is employed to capture temporal features in the physical domain, and the DNN is dedicated to capturing statistical features in the cyber domain. Moreover, a hierarchical detection and dynamic thresholding mechanism is presented to compensate for the poor performance of traditional distributed agents in microgrid environments. The experimental results and analysis demonstrate that the FDIA detection accuracy of the proposed scheme is better than several other benchmark detectors, and verify the effectiveness of the proposed scheme in cross-domain information detection.

As communication networks and the Internet of Things (IoT) converge in water critical infrastructure, it facilitates smart online monitoring of water supply systems, but it also significantly increases the incidence of cyber attacks. Additionally, it becomes especially challenging to detect attacks timely as they become more sophisticated and multifaceted. To address this issue, we propose a stacking ensemble network detection model based on industrial processes in water treatment systems, where the attacks are identified through recognizing anomalies caused by the underlying dynamics of the industrial control system. In the proposed detection model, we utilize a multilayer perceptron (MLP) to facilitate multi-module training, taking into consideration the multi-stage process of the water conservancy system. Furthermore, we adopt Long Short-Term Memory Recurrent Neural Network (LSTM) to effectively integrate the predicted results. To optimize the performance of this model, we leverage a Population Based Training (PBT) of Neural Networks to finetune the parameters. For each sensor parameter, the residuals between the predicted and measured data are classified using a sliding window-based dynamic threshold to identify anomalies. Based on the Real Safe Water Treatment (SWaT) system dataset, we evaluate the accuracy of our proposed scheme with three other schemes. The experimental comparison results show that the accuracy of our scheme is better than those of all the other three schemes.

As traditional automobiles evolve into the Internet of Vehicles (IoV), the increasingly frequent interactions between intelligent vehicles and external environments make cybersecurity a critical issue. While existing machine learning-based automotive intrusion detection methods demonstrate strong detection performance, most rely on supervised learning frameworks. These approaches not only incur high manual data labeling costs but also show significant limitations when handling real-world, unlabeled attack samples. In this paper, we propose an efficient intrusion detection system for in-vehicle networks based on self-supervised contrastive learning. By leveraging data augmentation, we construct positive sample pairs and learn robust feature representations through joint training using both reconstruction loss and contrastive loss. Additionally, we fuse the features extracted from the backbone and core networks into global representations for downstream classification tasks. Experiments on real in-vehicle intrusion datasets show that SSC-IDS achieves strong performance in anomaly detection. Furthermore, we test the model’s robustness under varying rates of anomaly contamination.

With the explosive growth of mobile devices and communication facilities, electromagnetic interference (EMI) has become a common phenomenon affecting the communication band. Based on the shielding capability of electromagnetic bands in EMI, frequency selective surfaces (FSSs) are used to shield or suppress specific electromagnetic bands. Additionally, EMI can be negative control and may change the EMI band. Thus, a single FSS cannot effectively shield EMI due to its limited shielding capacity. To address this issue, we first construct a novel interference shielding model to guard the target area. The related shielding problem is modeled as the UAV-driven FSS (UFSS) configuration problem. Second, we propose an intelligent configuration method based on a stochastic game to solve the configuration optimization problem effectively. In the proposed method, we model the interaction between UFSSs and interferers as a stochastic game, where we provide each UFSS with two different options for updating its shielding configuration strategy. According to the shielding configuration strategy generated by the proposed stochastic game, we propose a square loop resource allocation model based on resource constraints to promote each UFSS to update its square loop. Finally, the numerical results and analysis show that our proposed method is more effective and feasible than other band shielding configuration schemes.

To decrease the loss of stealing electricity on the smart grid, power theft detection (ETD) is widely concerned and studied. Due to missing values, invalid values, and non-linear data relationships in the grid data, existing electricity theft detection methods will generally lead to training errors, non-convergence, and low accuracy. To address those issues, we propose a CNN-BiLSTM-Attention method to detect electricity stealing behavior. Since the number of normal users in the grid is much larger than that of abnormal users, we propose an improved WGAN-GP (LWGAN-GP) algorithm to solve the data imbalance problem. Based on the generated balanced dataset, CNN and BiLSTM are adopted to extract features from current data and historical data, respectively. To further improve the classification accuracy, we use an attention mechanism to assign weights to the relevant features. Moreover, experiments based on real datasets show that the ACC, AUC, and F1 of the proposed CNN-BiLSTM-Attention method are greater than those of the other three machine learning and deep learning models.