Federated Learning (FL) enables collaborative model training without centralized data sharing, thereby preserving user privacy. However, the decentralized nature of FL exposes the system to Byzantine clients that can maliciously manipulate local model updates using attacks such as random noise injection, sign flipping, or model poisoning. Effective Byzantine detection is therefore essential to ensure the robustness of the global model. In this work, we propose a Hierarchical Handshake-based Federated Learning (HH-FL) framework. The system organizes clients into a tree-structured hierarchy, where each cohort performs local training and securely aggregates updates before forwarding them to the parent node. Within each cohort, clients engage in a handshake-based similarity check, computing pairwise distances between model updates. Leveraging this mechanism, we incorporate proven Byzantine-resilient aggregation methods such as Krum and Multi-Krum. Additionally, we introduce two novel detection schemes: Binary Segregation and a Continuous Rank-based Voting Algorithm, designed to dynamically identify and filter malicious clients. We evaluate and compare these defense strategies under multiple adversarial attack scenarios and experimental settings. Results demonstrate that our hierarchical approach enhances detection capability and improves global model robustness against Byzantine threats in Federated Learning.
#Research Paper tool.