HisHFL: Enhancing Hierarchical Federated Learning Security through Historical Knowledge

Federated Learning (FL) is a widely used distributed learning framework that aims to train global models while preserving client data privacy. However, with the proliferation of 6G/5G systems, many existing FL systems face challenges with frequent model exchanges and large-scale parameter communication between clients and cloud servers, resulting in high communication overhead and delays. To mitigate this, the Hierarchical Federated Learning (HFL) framework leverages mobile edge computing to aggregate client edge models, reducing transmission traffic. Despite its advantages, hierarchical FL is vulnerable to security threats, especially when attackers submit malicious model updates to higher layers, executing poisoning attacks that disrupt the training process. Since the aggregator lacks direct access to the original data, hierarchical FL is susceptible to such attacks. To address this, we propose a fine-grained, history-based Hierarchical Federated Learning framework-HisHFL, designed to effectively defend against poisoning attacks. HisHFL utilizes a dual-layer robust secure aggregation algorithm based on historical information to distinguish malicious attackers from benign participants, thereby protecting the hierarchical FL system from the impact of poisoning attacks. Experimental results demonstrate that HisHFL significantly improves the global model accuracy under application model and data poisoning attacks, outperforming existing state-of-the-art methods.