tensorflow/privacy

Provides differentially private TensorFlow optimizers (SGD, Adam, etc.) with per-example gradient clipping and built-in privacy accounting tools to measure epsilon/delta guarantees. Includes optimized Keras implementations for Dense and Embedding layers that achieve DP training without meaningful performance overhead by leveraging fast per-example gradient computation. Integrates directly with TensorFlow 2.x and tf.keras, with tutorials demonstrating how to wrap standard optimizers into privacy-preserving variants.

2,003 stars. Actively maintained with 3 commits in the last 30 days.