We study the problem of differentially private stochastic convex optimization (DP-SCO) with heavy-tailed gradients, where we assume a \(k^{\text{th}}\)-moment bound on the Lipschitz constants of sample functions rather than a uniform bound. We propose a new reduction-based approach that enables us to obtain the first optimal rates (up to logarithmi...

Private Stochastic Convex Optimization with Heavy Tails: Near-Optimality from Simple Reductions