Abstract Federated learning (FL) is a distributed machine learning framework that coordinates clients to train models on their private datasets via a centralized server, thereby mitigating data privacy risks. However, the communication channels involved in this process are untrusted, leaving FL vulnerable to model poisoning attacks launched by adversaries through man-in-the-middle techniques. Such attacks can degrade the accuracy of the global model and ultimately cause the entire FL training process to fail. In this paper, we propose a defense mechanism that integrates secure verification with watermarking, with the primary goal of ensuring the integrity of models transmitted over communication channels and enabling highly reliable FL deployment. Our mechanism leverages a dual watermarking method: first, models are marked using specially generated samples, and then these samples are further watermarked based on a class histogram-inspired approach. This dual strategy enhances both model detection and watermark stealthiness. The key innovation of our method lies in its sensitivity to subtle tampering while imposing no loss in model accuracy. Experimental results demonstrate that our defense mechanism significantly strengthens the resilience of FL models against sophisticated model poisoning attacks, while maintaining high accuracy and reliability.

Clustered Federated Learning (CFL) has emerged as an effective solution to address data heterogeneity in traditional Federated Learning (FL). However, the intrinsic cluster-based structure of CFL introduces new privacy risks, making it more vulnerable to client-level inference attacks. In this paper, we propose a novel client-level data reconstruction attack based on Conditional Generative Adversarial Networks (cGANs), which exploits intra-cluster similarities to enhance the quality of reconstructed private data. Unlike prior works, our attack requires only partial access to a victim’s model updates through passive eavesdropping, thereby reflecting a more realistic threat model in decentralized and resource-constrained environments such as the Internet of Things (IoT). To mitigate this threat, we develop a lightweight and adaptive defense mechanism grounded in Local Differential Privacy (LDP). Our design incorporates dynamic privacy budget decay, selective layer-wise noise injection, and real-time similarity-guided adaptation. This approach achieves a favorable privacy-utility trade-off while explicitly addressing the computational, communication, and latency constraints inherent in IoT environments. Experimental results demonstrate that our proposed attack improves reconstruction similarity by up to 20% compared with existing baselines, while the defense reduces attack success rate by 27.2% with only a 3.3% accuracy drop. Moreover, it significantly lowers computational cost—reducing FLOPs by 42.7%, memory usage by 23.4%, and DP noise processing time by 45.5%—without introducing additional communication overhead. These findings highlight the underestimated privacy vulnerabilities in CFL and underscore the necessity of efficient, context-aware defense strategies.

This decade witnesses the privilege of deep learning in intelligent fault diagnosis. However, the limited interpretability poses significant challenges to comprehending and trusting the decision-making mechanisms. To tackle these issues, this paper proposes a knowledge distillation framework that synthesizes dynamic and causal a priori, aiming to capture the fault mechanisms while diminishing confounding side-effect embedded in data. Firstly, this study models rolling bearings using dynamic a priori knowledge, and the simulated data imbued with fault dynamics are then fed into the teacher model training pipeline to yield a deeper understanding of mechanical failures. Secondly, in cope with the probability drift and inconsistency in data collected from variable operating conditions, features undergo weighted fusion according to the causal a priori amongst variables to avoid “correlation trap” in fault classification. Finally, the knowledge distillation module enioys the above two steps adjusting parameters according to both priori simultaneously to achieve a better interpretability in fault diagnosis. In the experiment, the study demonstrates that the proposed framework enhances both the accuracy and interpretability of the model, achieving a remarkable accuracy rate of 95.8% under variable operating conditions.

In modern manufacturing, significant advances in rolling bearing fault diagnosis have been driven by artificial intelligence. However, current intelligent fault diagnosis frameworks struggle to effectively capture causal relationships between time-series data and labels within multicause systems (i.e., complex industrial systems where equipment failures arise from the interplay of multiple coupled causes rather than a single factor), as well as represent the varying significance of features throughout fault progression. These issues lead to a considerable decrease in the diagnostic effectiveness of the model. To address these issues, a causal discovery method combined with a feature weighting strategy is proposed to enhance the model’s generalizability. Specifically, this article proposes the information flow and weighting for domain adaptation framework. First, this study uses information flow techniques to quantify the transfer of information between time-series data variables, thereby determining the causal direction among them. Second, a weighting strategy is developed based on the causal force between different causes. This approach enhances the model’s ability to accurately identify fault characteristics. Finally, data processing and feature weighting modules are integrated into practical applications, resulting in a comprehensive fault diagnosis framework powered by adversarial neural networks. Experimental results demonstrate that this method significantly enhances the fault recognition rate and generalization performance of the model.

Few-shot learning (FSL) techniques have been introduced to address the challenge of limited datasets in mechanical fault diagnosis. However, most existing FSL methods primarily focus on input–output correlations and neglect causal relationships, which limits the interpretability and robustness of diagnostic results. To tackle this issue, we propose a causal disentanglement few-shot relation metric network for mechanical fault diagnosis, comprising feature encoding, causal intervention, causal disentanglement, and relation metric modules. The causal intervention module performs linear interpolation on amplitude information (encoding low-level statistics) while preserving phase information (encoding high-level semantics) to intervene causally on the frequency-domain image. Fault features are extracted via the feature encoder module, and a factor disentanglement loss in the causal disentanglement module transforms them into independent causal features with explicit causal relationships. The relation metric module learns pairwise causal feature distances through meta-task training, thus constructing a trainable similarity metric space. This approach can effectively capture the differences in causal fault features between samples, enhancing the interpretability and generalization ability of the model. Experiments on both public and laboratory datasets demonstrate superior performance over state-of-the-art methods.

Abstract Federated learning (FL), celebrated for its privacy‐preserving features, has been revealed by recent studies to harbour security vulnerabilities that jeopardize client privacy, particularly through data reconstruction attacks that enable adversaries to recover original client data. This study introduces a client‐level handwriting forgery attack method for FL based on generative adversarial networks (GANs), which reveals security vulnerabilities existing in FL systems. It should be stressed that this research is purely for academic purposes, aiming to raise concerns about privacy protection and data security, and does not encourage illegal activities. Our novel methodology assumes an adversarial scenario wherein adversaries intercept a fraction of parameter updates via victim clients’ wireless communication channels, then use this information to train GAN for data recovery. Finally, the purpose of handwriting imitation is achieved. To rigorously assess and validate our methodology, experiments were conducted using a bespoke Chinese digit dataset, facilitating in‐depth analysis and robust verification of results. Our experimental findings demonstrated enhanced data recovery effectiveness, a client‐level attack and greater versatility compared to prior art. Notably, our method maintained high attack performance even with a streamlined GAN design, yielding increased precision and significantly faster execution times compared to standard methods. Specifically, our experimental numerical results revealed a substantial boost in reconstruction accuracy by 16.7%, coupled with a 51.9% decrease in computational time compared to the latest similar techniques. Furthermore, tests on a simplified version of our GAN exhibited an average 10% enhancement in accuracy, alongside a remarkable 70% reduction in time consumption. By surmounting the limitations of previous work, this study fills crucial gaps and affirms the effectiveness of our approach in achieving high‐accuracy client‐level data reconstruction within the FL context, thereby stimulating further exploration into FL security measures.

Edge computing (EC) provides an efficient and low-latency computing architecture for mobile multimedia communications. Blockchain-enabled EC can offer enhanced security and data privacy protection in the system, whereas throughput remains a big concern for the blockchain. Sharding is a promising solution to increase the throughput at the cost of complex cross-shard transaction verification. The popular two-phase commit protocol (2PC) can ensure the consistency of cross-shard transaction processing. However, in the existing schemes based on 2PC, the number of intra-shard consensus invocations is proportional to the number of transactions, which imposes a great challenge on the system throughput and adaptivity improvement in sharding blockchains under dynamic transaction processing demands and capacities. In this article, we propose a transaction processing scheme based on 2PC, such that multiple transactions can be simultaneously processed in a batch during every execution of the consensus. Furthermore, we model the problem of transaction allocation to batches as a communication load balancing problem, aiming to balance the inter-shard communications within each batch under the shard processing capacity constraint. We also propose an effective Batch Transaction Processing algorithm (BTP) for the problem. Theoretical analysis proves that BTP is a 3-approximation algorithm for the communication load balancing problem. In the simulations and experiments on BlockEmulator, BTP respectively improves the system throughput and total transaction processing time by at least 29.41% and 22.64% over the state-of-the-art cross-shard transaction processing schemes, which demonstrates the superior adaptivity performance of BTP.

As the popularity of automobiles increases, major cities around the world are facing serious problems of urban traffic congestion, and the increasingly severe traffic congestion will limit the development of cities. For urban road networks, relieving congestion at traffic bottlenecks can effectively improve overall traffic congestion, so identifying traffic bottlenecks in the road network is essential. Some researchers currently use spatial features based on section congestion propagation to predict bottlenecks, but their temporal features are also an important factor. A new bottleneck definition that considers both the temporal and spatial characteristics of congestion propagation is proposed. A congestion propagation dynamic model is built based on this definition, and the congestion propagation probability between different time periods is estimated using causal anonymous walks. The total congestion cost of road segments is then calculated, and traffic bottlenecks in the road network at different time periods are predicted. Finally, we used SUMO to simulate the Sioux city road network and collected 30 days of traffic flow information to construct a congestion propagation dataset. We then trained a causal anonymized walking network to estimate the congestion propagation probability for different time periods and predict traffic bottlenecks. By alleviating congestion on bottleneck sections, we demonstrated the effectiveness of this method in predicting bottlenecks and improving traffic flow for the road network as a whole.

In recent years, several few-shot learning (FSL) approaches for industrial equipment fault diagnosis have emerged to tackle the challenges posed by small fault diagnosis datasets. However, the existing FSL approaches model the correlation between input and output variables while ignoring causality, which cannot ensure that the diagnosis results are interpretable and robust. To tackle this problem, this article introduces a causal intervention relation network for cross-component few-shot fault diagnosis from the causal perspective. The model comprises a feature encoding module, a causal intervention module, and a relation measure module. The feature encoding module and the relation measure module establish a trainable similarity metric space through the training of multiple metatasks, where they learn the feature distances between sample pairs. Importantly, in causal intervention module, we model the causal structure of the metalearning process of few-shot fault diagnosis to find the causal fault features and the confounder factor, i.e., the metatraining diagnosis knowledge. Correspondingly a backdoor adjustment approach via a combination of class-based adjustment and feature adjustment is designed to realize the causal calibration of the few-shot fault diagnosis model. In such way, the model can capture causal invariant features between various components with significant distributional differences, thus enhancing the model's interpretability and its capacity for generalization. We perform experiments on two openly accessible datasets and a dataset constructed in our laboratory. The experimental results demonstrate that the model outperforms existing state-of-the-art approaches.

Abstract The Energy Harvesting Roadside Unit (EH‐RSU) with self‐powered module will not only effectively reduce the communication load of regional Vehicular Ad Hoc Networks, but also enjoys a low deployment cost. Given the imbalance in communication demands invoked by transportation systems, the EH‐RSU should allocate energy appropriately in accordance with its energy harvesting rate to ensure the communication safety of vehicles within its coverage. Firstly, we propose a novel attention‐based spatial‐temporal graph convolutional network (ASTGCN) to predict the communication load around the EH‐RSU in the road network through the surrounding vehicle information. Secondly, we use the predicted communication load as part of the input parameters to neural network and leverage a double deep Q network to ameliorate the operating states switching strategy of EH‐RSUs by reinforcement learning so that they achieve a more satisfying effective time with limited resources. Finally, we built a dataset by simulation to validate the effectiveness of our model. The results show that our prediction model has a better accuracy and the improved strategy has higher efficiency compared with other methods.

Due to less degradation data and the inconsistent data distribution of different bearings, remaining useful life (RUL) prediction methods based on deep learning still do not yield satisfactory predictive results. Using RUL prediction model trained with one bearing sample but tested with another bearing sample is challenging. To solve this problem, in this paper a new deep transfer learning-based RUL prediction method (DTL-RULPM) is proposed. We adopt min-max normalization to normalize the original vibration data of bearing. A three-layer sparse autoencoder is designed to extract the deep features of the source domain. Random data with standard normal distribution is generated with the consistent dimension of the high-dimensional features of the source domain. Maximum mean discrepancy (MMD) is used to minimize the probability distribution distance between the features of the source domain and the randomly generated data, such that the model can learn domain-invariant features of different bearings. Then we adopt a bi-directional long and short-term memory (Bi-LSTM) network to predict the RUL of the bearing. We use the IEEE PHM Challenge 2012 dataset to verify the proposed method. The results demonstrate that the proposed method improves the RUL prediction accuracy and robustness of different bearings.

The compound fault signal of bearings is coupled and complex, thereby compound fault diagnosis is a difficult problem in bearing fault diagnosis. The existing deep learning models can extract fault features when there are a large number of labeled compound fault samples. In the industrial scenarios, collecting and labeling sufficient compound fault samples are unpractical. Using the model trained on single fault sample to identify unknown compound fault is challenging and innovative. To address this problem, we propose a Zero-shot Learning Compound Fault Diagnosis Model of bearing (ZLCFDM). First, we design a semantic encoding method to express the semantic vectors of single fault and compound fault according to the fault characteristics. Second, a convolutional neural network is designed to extract the time-frequency visual features of compound fault signal. Then we embed the semantic vector of the fault into the visual space of the fault data. The cosine distance is merged into K-nearest neighbor (KNN) to measure the distance between the visual features and the semantic vectors of the compound faults, such that the model can identify the categories of unknown compound faults. To validate the proposed method, we conduct experiments on self-built testbed. The results demonstrate that the identification accuracy of compound fault can reach 77.73% when the model trained without any compound fault samples. This is the first time to propose the compound fault diagnosis of bearing base on zero-shot learning.

Due to the increasing usage of cloud computing applications, it is important to minimize energy cost consumed by a data center, and simultaneously, to improve quality of service via data center management. One promising approach is to switch some servers in a data center to the idle mode for saving energy while to keep a suitable number of servers in the active mode for providing timely service. In this paper, we design both online and offline algorithms for this problem. For the offline algorithm, we formulate data center management as a cost minimization problem by considering energy cost, delay cost (to measure service quality), and switching cost (to change servers’s active/idle mode). Then, we analyze certain properties of an optimal solution which lead to a dynamic programming based algorithm. Moreover, by revising the solution procedure, we successfully eliminate the recursive procedure and achieve an optimal offline algorithm with a polynomial complexity. For the online algorithm, We design it by considering the worst case scenario for future workload. In simulation, we show this online algorithm can always provide near-optimal solutions.