Differential Privacy for Federated Learning with Dynamic Group Membership

Abstract

Federated learning (FL) offers a promising approach to machine learning by enabling collaborative model training across decentralized devices without directly exchanging data. However, traditional FL methods often struggle to provide strong differential privacy guarantees, especially when group membership and data heterogeneity change dynamically. This paper proposes a novel framework for enhancing differential privacy in FL by incorporating a dynamic privacy budget allocation scheme. The core idea is to adapt the privacy level based on the size and diversity of each participating group, mitigating the impact of fluctuating group memberships and varying data distributions. We present a detailed analysis of the challenges posed by dynamic group membership and demonstrate the effectiveness of our approach through a theoretical framework. The proposed method aims to strike a balance between model accuracy and privacy preservation, contributing to a more robust and reliable FL system. The key contributions lie in the adaptive privacy allocation and the addressing of non-static group dynamics.