FedPriv: Personalized Differential Privacy in Federated Learning for Heterogeneous Data

With the widespread adoption of interactive, mobile,and wearable technologies, an increasing amount of privacysensitive and statistically heterogeneous sensor data is beingcontinuously generated by ubiquitous user devices. Federatedlearning (FL) enables multiple parties to jointly train a machinelearning model while keeping their raw data private. Nonetheless,under realistic conditions characterized by pronounced statisticalheterogeneity across client distributions, traditional FL methodssuffer from degraded performance. Moreover, uniform privacymechanisms fail to accommodate the diverse privacy needs ofindividual users. To this end, we propose a privacy-preserving FLframework, FedPriv, for heterogeneous mobile sensor data, aiming to improve the performance of the model in a heterogeneousdata environment while protecting user privacy. Specifically,FedPriv is built upon two core components that we design: (1)a novel aggregation mechanism that dynamically adjusts eachclient’s contribution according to the discrepancy between localand global data distributions, thereby mitigating adverse effectsof data heterogeneity on the model performance; and (2) apersonalized differential privacy strategy that allocates privacybudgets adaptively, depending on each client’s data sensitivity,enabling fine-grained privacy protection without compromisingutility. In addition, FedPriv can be seamlessly integrated asa plug-and-play module into existing FL pipelines, enhancingmodel performance without requiring architectural modifications. Security analysis demonstrates that FedPriv satisfies formalϵ-differential privacy guarantees while supporting personalizedprivacy levels across users. Extensive experiments on six realworld datasets, consisting of four image datasets and two humanactivity recognition datasets, show that FedPriv achieves a 2.75%-5% improvement in model performance and reduces convergencetime by an average of 15% when compared to existing baselinesunder heterogeneous data settings.