p_exponential_mechanism

We propose a mechanism family (p-power exponential mechanism family) of DP to improve the performance of the Gaussian mechanism in machine learning.

TensorFlow Privacy

This folder contains the source code for TensorFlow Privacy, which is a Python library developed by google (https://github.com/tensorflow/privacy) for training machine learning models with differential privacy. The detailed procedures for setting up TensorFlow Privacy in our repository are as follows.

Installing TensorFlow Privacy

First, clone this GitHub repository into a directory of your choice:

git clone https://github.com/private-mechanism/p_exponential_mechanism

You can then install the local package in "editable" mode in order to add it to your PYTHONPATH:

cd P_Exponential_Mechanism

pip install -e .

Comparing to the original version of TensorFlow Privacy, we have added one python file under privacy/optimizers and privacy/dp_query folders respectively.

In dp_fixedvariance_query.py under privacy/dp_query folder, we present how to sample noises from the p-exponential distributions.

In flatten_optimizer.py under privacy/optimizers folder, we present how to generate the DP optimizer under the p-exponential mechanism.

Tutorials directory

The tutorials/ folder contains the scripts presenting how to apply p-exponential mechanism in the machine learning models. In particular, we apply p-exponential mechanism into Logistic Regression(LR) on mnist dataset, and Conventional Neural Networks(CNN) on both mnist and cifar10 datasets. Additionally, we also present the scripts for computing the total privacy loss using moments accountant technique in p-exponential mechanism.

Contacts

If you have any questions that cannot be addressed by raising an issue, feel free to contact:

• zfy1454236335 (@stu.xjtu.edu.cn)