From Privacy to Generalization: Linear Max-Information Bounds for DP-SGD

In three linesTheoretical paper proving a finite-sample bound on approximate max-information of DP-SGD with linear scaling in dataset size. Derives a general PAC-Bayes generalization bound where the prior distribution is learned by DP-SGD, and a generalization bound for DP-SGD-trained models with complexity term explicitly controlled by optimization hyperparameters.

Summary generated by Claude — human-verified