Filtering Method Against Malicious Clients by Edge-Based Accuracy Validation in Hierarchical Federated Learning

In modern distributed learning environments, Hierarchical Federated Learning (HierFL) offers both scalability and communication efficiency within a server–edge–client architecture. However, malicious clients, such as those performing label flipping, can significantly degrade the performance of the global model in HierFL. Additionally, existing aggregation methods often fail to defend against such attacks effectively. To address this issue, this paper proposes a filtering approach that verifies accuracy at the edges. The proposed method evaluates local models submitted by each client using an independent verification dataset at the edges. It transmits only updates that meet a specific accuracy threshold to the server. This approach demonstrates that even as the proportion of malicious clients increases, the proposed method can reliably maintain the accuracy of the global model compared to existing average-based aggregation methods. This proves that the proposed method can enhance security and reliability in hierarchical federated learning environments.