Differentially Private Deep Learning

This project focuses on applying Differential Privacy to deep learning. The primary model used here is ResNet-20, although testing was also done on WideResNet-16-4. For parallelization, Differentially Private Distributed Data Parallel (DPDDP) is used. We have also implemented the Differentially Private Importance Sampling algorithm from the DPIS paper by Wei, et al.