Differential Privacy for Federated Learning with Non-IID Data

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without directly exchanging data. However, the inherent non-IID (independent and identically distributed) nature of data across clients poses a significant threat to the privacy guarantees provided by traditional differential privacy (DP) mechanisms. This paper introduces a novel differential privacy framework specifically designed for federated learning scenarios with non-IID data. Our approach incorporates adaptive noise scaling, dynamically adjusting the privacy budget based on the measured data heterogeneity among clients. This tailored mechanism mitigates the increased privacy risk associated with non-IID data, providing a more robust and effective solution for privacy-preserving federated learning. We demonstrate the effectiveness of our method through theoretical analysis and highlight its advantages over conventional DP techniques in this context.