Distributed Differential Privacy with Dynamic Grouping for Federated Learning

Abstract

Federated learning (FL) offers a promising approach to train machine learning models on decentralized data sources without directly exchanging the data itself. However, the inherent privacy risks associated with FL remain a significant concern. Traditional differential privacy (DP) techniques often struggle to effectively protect privacy in FL settings, particularly when dealing with heterogeneous data distributions across participating devices. This paper proposes a novel framework for distributed differential privacy in federated learning that incorporates dynamic grouping of devices based on their data similarity and sensitivity. Our approach continuously analyzes data distributions across participating devices and adjusts grouping assignments to mitigate information leakage. We aim to achieve stronger differential privacy guarantees while maintaining reasonable data utility. The core of our method lies in a dynamic grouping algorithm which optimizes for both privacy and utility. We formally define the privacy loss and utility metrics, and provide a theoretical analysis of our framework. The key innovation is the adaptive grouping strategy, leading to more effective privacy protection.