In 📔 this notebook we demonstrate how powerful it is to use JAX based DP-SGD along with the dp-accounting library to train deep learning models with Differential Privacy (DP). We show how to use these tools on a small toy problem: histogram estimation, and compare this generic approach to some well known benchmarks.
Along the way we get to discover some of the problems one is commonly faced with when tuning the hyper-parameters of a DP-SGD, in particular its clipping threshold.
If you own privacy-sensitive data and would like more people to use this data without having to worry about privacy risk, you might be interested by what Sarus Technologies does.
.png)
