Lossy Compression of Noisy Data for Private and Data-Efficient Learning

Storage-efficient privacy-preserving learning is crucial due to increasing amounts of sensitive user data required for modern tasks. We propose a framework reducing the storage cost while at same time providing privacy guarantees, without essential loss in utility learning. Our method comprises noise injection followed by lossy compression. show that, when appropriately matching compression distribution added noise, compressed examples converge, distribution, that noise-free training as sample size (or dimension data) increases. In this sense, essentially maintained, and leakage quantifiable amounts. present experimental results on CelebA dataset gender classification find our suggested pipeline delivers practice promise theory: individuals images are unrecognizable less recognizable, depending level), overall substantially reduced, with no (and some cases slight boost) accuracy. As an bonus, experiments suggest yields substantial boost robustness face adversarial test data.