We revisit the issue of safe aggregation of high-dimensional vectors in a two-server system comparable to Prio. These techniques are usually used to combination vectors comparable to gradients in personal federated studying, the place the mixture itself is protected by way of noise addition to make sure differential privateness. Present approaches require communication scaling with the dimensionality, and thus restrict the dimensionality of vectors one can effectively course of on this setup.
We suggest PREAMBLE: {bf Pr}ivate {bf E}fficient {bf A}ggregation {bf M}echanism by way of {bf BL}ock-sparse {bf E}uclidean Vectors. PREAMBLE builds on an extension of distributed level capabilities that permits communication- and computation-efficient aggregation of {em block-sparse vectors}, that are sparse vectors the place the non-zero entries happen in a small variety of clusters of consecutive coordinates. We present that these block-sparse DPFs may be mixed with random sampling and privateness amplification by sampling outcomes, to permit asymptotically optimum privacy-utility trade-offs for vector aggregation, at a fraction of the communication price. When coupled with latest advances in numerical privateness accounting, our method incurs a negligible overhead in noise variance, in comparison with the Gaussian mechanism used with Prio.
- ** Work finished whereas at Apple
- † Aarhus College
