Trade-off Functions for DP-SGD with Subsampling based on Random Allocation: Tight Upper and Lower Bounds
The paper analyzes Differential Privacy Stochastic Gradient Descent (DP-SGD) using random shuffling to find tight bounds for its trade-off function, which balances privacy and model utility. It establishes specific noise regimes where meaningful differential privacy can be guaranteed, providing concrete parameter recommendations for single-epoch training in federated learning.