Differentially private machine learning at scale with JAX-Privacy
Google Research
Google released JAX-Privacy 1.0, a toolkit built on the JAX numerical computing library that enables researchers and developers to implement differentially private machine learning algorithms at scale. The library provides core components for differential privacy including per-example gradient clipping, noise addition, and auditing tools, with support for training large language models like VaultGemma through JAX's native parallelism features. The open-source release aims to lower barriers for building privacy-preserving AI applications by integrating differential privacy into modern machine learning workflows.
Why it matters
Algorithms & Theory