User-level Differentially Private Stochastic Convex Optimization: Efficient Algorithms with Optimal Rates

We examine differentially personal stochastic convex optimization (DP-SCO) underneath user-level privateness the place every consumer could maintain a number of information gadgets. Present work for user-level DP-SCO both requires super-polynomial runtime or requires variety of customers that grows polynomially with the dimensionality of the issue. We develop new algorithms for user-level DP-SCO that acquire optimum charges, run in polynomial time, and require plenty of customers that develop logarithmically within the dimension. Furthermore, our algorithms are the primary to acquire optimum charges for non-smooth capabilities in polynomial time. These algorithms are primarily based on multiple-pass DP-SGD, mixed with a novel personal imply estimation process for concentrated information, which applies an outlier elimination step earlier than estimating the imply of the gradients.