JAX-Privacy Benchmarking Suite

A standalone, easily forkable benchmarking suite developed as a new top-level directory (benchmarks/) within the Google DeepMind JAX-Privacy repository, sitting alongside the existing jax_privacy/ and examples/ directories. The suite enables researchers and developers to evaluate differentially private machine learning mechanisms across standardised models and datasets using JAX-native components. It includes three model implementations (CNN, Transformer, and LoRA adapter for Gemma), eight benchmark datasets organised by differential privacy unit (example-level: CIFAR-10, IMDB; user-level: StackOverflow, FLAIR, CelebA; multi-owner: Enron, ArXiv), privacy accounting and metrics utilities, a unified training script with CLI support, and a comprehensive test suite. All implementations adhere to JAX-Privacy code conventions, utilising core components such as clipped_grad, execution_plan, and batch_selection strategies. The project also includes beginner-friendly documentation published via ReadTheDocs. Developed as a project under Google DeepMind.