Enhanced Spectral Clustering Robust Aggregation for Lens Detection in Federated Learning Against Byzantine Attacks

Although federated learning (FL) addresses the issues of centralized data storage and privacy leakage, its distributed nature makes it vulnerable to malicious clients. These malicious participants introduce malicious parameters during the training procedure, which can significantly impact the model's accuracy. Existing algorithms such as Krum and median defend against attacks by capturing low-order features of data and requiring prior data. However, these approaches struggle to counteract gradually evolving Byzantine attacks. Therefore, we propose an unsupervised approach based on an enhanced spectral clustering algorithm (SCA) to identify malicious updates. First, we design a method for constructing an undirected weighted graph using the Gaussian kernel function. This approach maps features among data into an infinite-dimensional Hilbert space, enabling better capture of high-order data features. Second, due to the similarity between Byzantine updates and benign updates in Euclidean space and cosine similarity scenarios, traditional robust aggregation algorithms fail to recognize them, causing models to fail to converge. To address this, a new lens detection method is designed. We calculate the Laplacian matrix through the undirected weighted graph and employ normalized cut (NCut) to partition the Laplacian matrix. This transforms the problem of identifying malicious clients into a graph partitioning problem. Furthermore, we conduct a convergence analysis of the proposed SCA. Experiments on four datasets under independent and identically distributed (IID) and non-IID (Non-IID) settings show this method has strong robustness to all tested attacks, while other defense methods cannot resist all of them.