A collaborative defense scheme against byzantine and evasion attacks for heterogeneous DFL

Decentralized Federated Learning (DFL) enhances privacy by processing data locally. However, its serverless architecture faces two major challenges: the severe data heterogeneity among clients, and the Byzantine during training phases and evasion attacks during inference phases. Existing solutions often address these issues in isolation. This paper proposes CentrifySim, a unified framework designed for Heterogeneous DFL that collaboratively defends against both attack vectors. It introduces three key innovations: The Feature Center Separation technique enhances the model's intrinsic robustness to adversarial perturbations through the enforcement of inter-class margin maximization and intra-class compactness constraints within the latent space. The Similarity-Driven Aggregation strategy significantly enhances model aggregation performance under non-independent and identically distributed (non-IID) data conditions by constructing an adaptive client similarity matrix. The Malicious Client Detection system effectively identifies and isolates malicious clients by integrating Principal Component Analysis (PCA) for dimensionality reduction and the Dirichlet Gaussian Mixture Model (DGMM), thereby ensuring the security and reliability of the aggregation process. These innovations have enhanced the robustness of the model. Comprehensive evaluations on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate that CentrifySim exhibits superior robustness under various attacks and data distributions. On the MNIST dataset, this approach can effectively defend against DBA attacks. It achieves a training accuracy of 97.73% while keeping the attack success rate at an extremely low level of 0.45%. Moreover, it demonstrates a consistent accuracy advantage in avoiding attacks during the inference stage.