Efficient Communication-Privacy Trade-offs in L2 Mean Estimation under Streaming Differential Privacy
This work introduces novel privacy accounting methods for the sparsified Gaussian mechanism that operate directly in L2 geometry, yielding mean square errors that converge fast to those of the uncompressed Gaussian mechanism. It also extends the sparsification scheme to the matrix factorization framework under streaming differential privacy, providing a precise accountant tailored for DP-FTRL type optimizers.