Federated Learning (FL) enables collaborative model training across untrusted devices while preserving data privacy. However, malicious participants can still launch attacks, especially poisoning attacks, during model aggregation. To defend against poisoning attacks, we propose a Coulomb Force-based Federated Learning (CFFL), a physics-inspired defense framework that integrates neighborhood Coulomb force into FL security mechanisms. CFFL effectively addresses detection failures in poisoning attacks arising from the “distance concentration” among high-dimensional data. Specifically, firstly, we establishes a local update model for clients, where the similarity between model updates is quantified through local Coulomb resultant force (LCRF), effectively distinguishing malicious and benign updates; Secondly, we develop ak-nearest neighbor (KNN)-based Coulomb resultant force anomaly detection (NCFAD) model that identifies malicious updates by isolating top-koutliers with largest neighborhood factor; Finally, experiment results validate that CFFL outperforms state-of-the-art (SOTA) baselines in defense performance and achieves model accuracy equivalent to the baseline FedAvg on NCI-1, PROTEINS_full, and MNIST datasets (90.82%, 92.23%, and 96.49%, respectively). This indicates that CFFL effectively mitigates the impact of poisoning attacks without compromising the benign performance of the aggregation model.

The rapid development of the Artificial Intelligence of Things (AIoT) opens up a new perspective for emerging service-based applications and becomes a major driver of diverse federated learning (FL) applications. However, due to the heterogeneity of nodes and the existence of free-rider attacks, some device nodes may launch free-rider attacks to obtain the global model without any contribution, which not only dampens the enthusiasm of legitimate participants but also undermines the fairness of node contributions. In this article, we propose a free-rider attack detection (FRAD) mechanism for FL with deep autoencoding Gaussian mixture model (DAGMM) based on contribution and reputation. Specifically, we first model the contribution values based on the computing resource, communication cost, and data quality of each device node. Then, based on PageRank algorithms, we design an optimal reputation-based model to fairly and precisely choose benign nodes to participate in federated training under information asymmetry. Furthermore, we develop the FRAD mechanism via DAGMM that combines historical contribution with reputation scores. Simulation results validate that the proposed mechanism in this article outperforms the state-of-the-art baselines$(\sim \!\! 2.14\times $and$\sim \!\! 1.1\times $on MNIST and CIFAR, respectively) in defending against free-rider attacks, and when the main clients are free-riders, i.e., 50% or even up to 80% of the free-riders, FRAD can still maintain a high defense performance against free-rider attacks.

Biomedical image segmentation of organs, tissues and lesions has gained increasing attention in clinical treatment planning and navigation, which involves the exploration of two-dimensional (2D) and three-dimensional (3D) contexts in the biomedical image. Compared to 2D methods, 3D methods pay more attention to inter-slice correlations, which offer additional spatial information for image segmentation. An organ or tumor has a 3D structure that can be observed from three directions. Previous studies focus only on the vertical axis, limiting the understanding of the relationship between a tumor and its surrounding tissues. Important information can also be obtained from sagittal and coronal axes. Therefore, spatial information of organs and tumors can be obtained from three directions, i.e. the sagittal, coronal and vertical axes, to understand better the invasion depth of tumor and its relationship with the surrounding tissues. Moreover, the edges of organs and tumors in biomedical image may be blurred. To address these problems, we propose a three-direction fusion volumetric segmentation (TFVS) model for segmenting 3D biomedical images from three perspectives in sagittal, coronal and transverse planes, respectively. We use the dataset of the liver task provided by the Medical Segmentation Decathlon challenge to train our model. The TFVS method demonstrates a competitive performance on the 3D-IRCADB dataset. In addition, the t-test and Wilcoxon signed-rank test are also performed to show the statistical significance of the improvement by the proposed method as compared with the baseline methods. The proposed method is expected to be beneficial in guiding and facilitating clinical diagnosis and treatment.

Abstract The wide application of Internet of Things technology has promoted the application of location‐based services (LBS). Users can enjoy various conveniences brought by LBS. However, mobile users should submit their location information to LBS which may lead to privacy disclosure. Therefore, users need to protect their privacy while enjoying the service. If the privacy protection problem cannot be solved, the development of mobile internet business will be greatly affected. The existing location privacy protection methods take little account of real‐time road conditions and are unable to resist link attacks. In this article, a k‐anonymous location privacy protection method based on Voronoi map was proposed. First, the Voronoi diagram was divided according to road network structure and real‐time road network data, and Voronoi unit was established with the intersection as the base point. Then, the base point is sent as an anchor instead of the user's actual location to the third‐party server. The Voronoi diagram generated by the k‐anonymity constituted an anonymous space to resist multi‐query attacks and protected the location. k‐1 user location points that are communicating were selected to form k‐anonymous set with the current user, and send the anonymous set to LBS. Theoretical analysis and experimental results show that the proposed method not only efficiently protected location privacy, but also provided high‐quality service.

In a recent study, it was shown that, with adversarial training of an attentive generative network, it is possible to convert a raindrop degraded image into a relatively clean one. However, in real world, raindrop appearance is not only formed by individual raindrops, but also by the distant raindrops accumulation and the atmospheric veiling, namely haze. Current methods are limited in extracting accurate features from a raindrop degraded image with background scene, the blurred raindrop regions, and the haze. In this paper, we propose a new model for an image corrupted by the raindrops and the haze, and introduce an integrated multi-task algorithm to address the joint raindrop and haze removal (JRHR) problem by combining an improved estimate of the atmospheric light, a modified transmission map, a generative adversarial network (GAN) and an optimized visual attention network. The proposed algorithm can extract more accurate features for both sky and non-sky regions. Experimental evaluation has been conducted to show that the proposed algorithm significantly outperforms state-of-the-art algorithms on both synthetic and real-world images in terms of both qualitative and quantitative measures.

An independent vector analysis (IVA) method base on variable-step gradient algorithm is proposed in this paper. According to the sEMG physiological properties, the IVA model is applied to the frequency-domain separation of convolutive sEMG mixtures to extract motor unit action potentials information of sEMG signals. The decomposition capability of proposed method is compared to the one of independent component analysis (ICA), and experimental results show the variable-step gradient IVA method outperforms ICA in blind separation of convolutive sEMG mixtures.