Cross-silo Federated Learning with Record-level Personalized Differential Privacy

Federated learning enhanced by differential privacy has emerged as a popularapproach to better safeguard the privacy of client-side data by protectingclients' contributions during the training process. Existing solutionstypically assume a uniform privacy budget for all records and provideone-size-fits-all solutions that may not be adequate to meet each record'sprivacy requirement. In this paper, we explore the uncharted territory ofcross-silo FL with record-level personalized differential privacy. We devise anovel framework named rPDP-FL, employing a two-stage hybrid sampling schemewith both client-level sampling and non-uniform record-level sampling toaccommodate varying privacy requirements. A critical and non-trivial problem isto select the ideal per-record sampling probability q given the personalizedprivacy budget ϵ. We introduce a versatile solution namedSimulation-CurveFitting, allowing us to uncover a significant insight into thenonlinear correlation between q and ϵ and derive an elegantmathematical model to tackle the problem. Our evaluation demonstrates that oursolution can provide significant performance gains over the baselines that donot consider personalized privacy preservation.