A Privacy-Preserving Decentralized Federated Learning Framework Based on Differential Privacy

Federated learning allows machine learning models to be trained by local clients without sharing data directly, providing a degree of privacy protection. However, existing studies reveal that traditional federated learning is prone to issues like gradient leakage and model poisoning, which can compromise user privacy and negatively impact the performance of the global model. In this paper, we propose a Decentralized Differential Privacy Federated Learning (DDPFL) framework that strengthens privacy protection in federated learning by integrating decentralized design with differential privacy. First, we incorporate differential privacy noise through the Laplace mechanism, applying noise via a hierarchical processing scheme. Additionally, we control the noise addition by employing an adaptive gradient pruning strategy. Then, we develop a decentralized federated learning framework that employs a committee scoring mechanism. The core of this mechanism is a scoring system that determines which clients are eligible to contribute to the aggregation of the global model. Theoretical analysis confirms that the proposed DDPFL framework meets differential privacy requirements, with the level of privacy protection adjustable by selecting different $\epsilon$ values. Experimental results show that DDPFL exhibits strong resilience to model poisoning attacks, and that performance degradation can be managed effectively by selecting an appropriate privacy protection budget.