Private Stochastic Convex Optimization: Optimal Rates in ℓ₁ Geometry

Stochastic convex optimization over an ℓ₁-bounded domain is ubiquitous in machine learning applications such as LASSO but remains poorly understood when learning with differential privacy. We show that, up to logarithmic factors the optimal excess population loss of any (ε, δ)-differentially private optimizer is plog(d)/n + √d/εn. The upper bound is based on a new algorithm that combines the iterative localization approach of Feldman et al. (2020a) with a new analysis of private regularized mirror descent. It applies to ℓ_p bounded domains for p ∈ [1, 2] and queries at most n^3/2 gradients improving over the best previously known algorithm for the ℓ₂ case which needs n² gradients. Further, we show that when the loss functions satisfy additional smoothness assumptions, the excess loss is upper bounded (up to logarithmic factors) by plog(d)/n + (log(d)/εn)^2/3. This bound is achieved by a new variance-reduced version of the Frank-Wolfe algorithm that requires just a single pass over the data. We also show that the lower bound in this case is the minimum of the two rates mentioned above.