Towards Privacy Preservation in Federated Learning: a Dynamic Multilayer Differential Privacy Framework

Federated learning (FL) is a distributed machine learning paradigm that addresses data privacy challenges in decentralized environments. Differential privacy (DP) is usually further used to FL for preventing model information leakage. However it often leads to a suboptimal trade-off between privacy and utility. To address this issue, this paper presents a novel approach to privacy protection in FL, introducing the Modular Multilayer Differentially Private Federated Learning (MMDPFL) framework. By integrating key features such as data entropy, gradient sensitivity, data variability, and mean square error between client predicted and actual values, MMDPFL dynamically adjusts privacy measures to align with each client's unique data characteristics. A privacy sensitivity score, derived through a weighted fusion of these metrics, categorizes clients into distinct privacy tiers (clusters), enabling tailored privacy budget allocation and noise injection strategies. Subsequently, the differential privacy protection level is adjusted according to the specific needs of each cluster. Experimental results show that MMDPFL improves accuracy by 2 %-4 % in 3, 5, and 8 clustering scenarios, while maintaining the same level of data privacy protection as the method with uniform privacy budget.