From personalized recommendations to scientific advances, AI models are improving lives and transforming industries. However, the impact and accuracy of these AI models is often determined by the quality of the data used. Large, high-quality datasets are essential for developing accurate and representative AI models, but they must be used in a way that protects individual privacy.
This is where JAX and JAX-Privacy come into play. Introduced in 2020, JAX is a high-performance numerical library designed for large-scale machine learning (ML). Core features such as automatic differentiation, just-in-time compilation, and seamless scaling across multiple accelerators make it an ideal platform for efficiently building and training complex models. JAX is a cornerstone for researchers and engineers pushing the limits of AI. Its surrounding ecosystem includes a robust set of domain-specific libraries, including Flax, which simplifies the implementation of neural network architectures, and Optax, which implements a state-of-the-art optimizer.
Built on JAX, JAX-Privacy is a robust toolkit for building and auditing differentially private models. It enables researchers and developers to quickly and efficiently implement differentially private (DP) algorithms for training deep learning models on large datasets, providing the core tools needed to integrate private training into modern distributed training workflows. The original version of JAX-Privacy was introduced in 2022 to allow external researchers to reproduce and validate some of the advances made regarding private training. Since then, it has evolved into a hub where research teams across Google integrate new research insights into DP training and auditing algorithms.
Today, we are proud to announce the release of JAX-Privacy 1.0. Integrating the latest research advances and redesigned with modularity in mind, this new version makes it easier than ever for researchers and developers to build DP training pipelines that combine cutting-edge DP algorithms with the scalability provided by JAX.
