Research on Differential Privacy Optimization Algorithm Based on DPSGD Model

This project studies a differential privacy stochastic gradient descent algorithm (DPSGD) based on MapReduce. The research direction of this article is to carry out research on efficient privacy protection issues for distributed data processing. Add Gaussian white noise pattern cropping to existing slope correction methods, and make some modifications to the gradient correction algorithm. A user information protection strategy based on user information has been proposed, and improvements have been made to the user information protection algorithm using this strategy. These improvements can map the privacy damage accumulation function to the sphere of the Frobenius module during the iteration process, which can improve its convergence performance.