The example code for the paper "Differentially Private Model Publishing for Deep Learning"
Some notes:
- The code was based on the original DP-SGD implementation in CCS'16 which was not maitained anymore. It requires Tensorflow 0.10.0. A fork canbe found here
- In the training code DumpzCDPAccountant is used and the real accounting is done during the training step. zCDPAccountant implementation is in privacy_accountant/accountant.py but not used yet.
Directory:
- The examples with decay functions are in dp_mnist and dp_cancerdataset
- The examples with validation based decay are in dp_mnist_validation and dp_cancerdataset_validation
- The privacy_accountant includes the code for comparing different privacy accountants, and conversion to epsilon, delta DP.