Differential Privacy for Federated Learning with Personalized Noise

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data, preserving data privacy. However, traditional differential privacy (DP) mechanisms often introduce significant noise into the global model updates, leading to substantial accuracy degradation. This paper presents a novel personalized differential privacy (P-DP) scheme for FL, where the noise level is dynamically adjusted based on the sensitivity of individual user data. We introduce a method for calculating per-user sensitivity values, considering local data distributions, and then employ adaptive noise scaling to minimize the privacy-utility trade-off. The proposed approach aims to achieve a better balance between privacy guarantees and model accuracy compared to standard DP methods in FL. The core contribution lies in the personalized allocation of the differential privacy budget, optimizing the system for a given application. Experimental results, although not presented here due to the focus on the method itself, would demonstrate the effectiveness of this approach.