Improving the Privacy-Utility Tradeoff in Differentially Private Machine Learning with Public Data

Explore two innovative algorithms, DP-RAFT and DOPE-SGD, designed to enhance the privacy-utility tradeoff in Differentially Private Machine Learning using public data. Learn how these techniques leverage available public information to improve model accuracy while maintaining strong privacy guarantees. Discover the benefits of DP-RAFT for scenarios with ample public data for pretraining, and understand how DOPE-SGD utilizes advanced data augmentation to maximize limited public data resources. Gain insights into overcoming the challenges of gradient clipping and noise addition in Differentially Private Stochastic Gradient Descent (DP-SGD) to achieve better model performance without compromising privacy protection.