FB open-source platform designed to help engineers with differential privacy
Facebook’s Opacus tool, offers new approach
Facebook wants to make it easier for engineers to use differential privacy in AI, with a new tool that limits impact on the primary dataset. Opacus, which trains PyTorch models with differential privacy, uses an algorithm that centers instead on intervening with parameter gradients. This follows recent announcements from Google to open source the differential privacy library, and Microsoft, which released WhiteNoise for Azure and GitHub.