FedHC: A Hierarchical Clustering Algorithm for Federated Anomaly Detection on Meta Computing

Currently,the research of federated learning(FL) in the field of anomaly detection is mainly based on the assumption that the data follows the principle of independent identically distribution(IID). On this basis, the accuracy of the detection task is improved by optimizing the model. However, in some real Meta Computing scenaros, data on edge devices are usually Non-IID, which leads the accuracy of trained models decreased significantly. In this study, we introduce the Federated Hierarchical Clustering algorithm(FedHC), which integrates hierarchical clustering mechanisms to reduce the inference of Non-IID in FL. The method achieve grouping based on client data similarity first, and trained each group independently. However,Relying solely on the data from a single group may limit the model's generalization capability. To address this issue, we further propose FedHCDW,a distance-weighted aggregation algorithm based on data distribution similarity, which can be effectively utilized by evaluating the similarity of data across different groups. The experimental results on the NSL-KDD dataset show that compared with FedAvg, FedHC not only achieves a significant improvement in accuracy of 1.7 %, but also accelerates the convergence speed of the model. Furthermore, compared with the FedHC algorithm, its improved version FedHCDW achieves an additional 1.1% increase in model accuracy, which proves the significant improvement of FedHCDW in model generalization ability.