FedPriv: Personalized Differential Privacy in Federated Learning for Heterogeneous Data

Abstract

With the widespread adoption of interactive, mobile, and wearable technologies, an increasing amount of privacy-sensitive and statistically heterogeneous sensor data is being continuously generated by ubiquitous user devices. Federated learning (FL) enables multiple parties to jointly train a machine-learning model while keeping their raw data private. Nonetheless, under realistic conditions characterized by pronounced statistical heterogeneity across client distributions, traditional FL methods suffer from degraded performance. Moreover, uniform privacy mechanisms fail to accommodate the diverse privacy needs of individual users. To this end, we propose a privacy-preserving FL framework, FedPriv, for heterogeneous mobile sensor data, aiming to improve the performance of the model in a heterogeneous data environment while protecting user privacy. Specifically, FedPriv is built upon two core components that we design: (1) a novel aggregation mechanism that dynamically adjusts each client's contribution according to the discrepancy between local and global data distributions, thereby mitigating adverse effects of data heterogeneity on the model performance; and (2) a personalized differential privacy strategy that allocates privacy budgets adaptively, depending on each client's data sensitivity, enabling fine-grained privacy protection without compromising utility. In addition, FedPriv can be seamlessly integrated as a plug-and-play module into existing FL pipelines, enhancing model performance without requiring architectural modifications. Security analysis demonstrates that FedPriv satisfies formal $\epsilon$-differential privacy guarantees while supporting personalized privacy levels across users. Extensive experiments on six real-world datasets, consisting of four image datasets and two human activity recognition datasets, show that FedPriv achieves a 2.75%-5% improvement in model performance and reduces convergence time by an average of 15% when compared to existing baselines under heterogeneous data settings.