privacy

Why we included this project

Training a model on data where a breach would be costly, whether customer records or medical logs, usually means giving up on the standard TensorFlow training loop. TensorFlow Privacy changes that by providing differentially private optimizers, DP-SGD variants of SGD and Adam, that drop into existing Keras models without rewriting the model code. The repo also ships analysis tools that estimate the privacy budget a training run actually consumes, plus tutorials on tuning the parameters. A fast per-example gradient clipping implementation, aimed at Dense and Embedding layers, removes most of the memory overhead and the need to fiddle with microbatch sizes, which makes private training feasible on ordinary hardware.