MLSec

It's a repository to implement some experiments on Machine Learning Security

Code Description

1. Data_Poisoning Directory

• label_poisoning_attack_svm.py

It aims to implement a data poisoning attack on labels with SVM classifier.

2. Membership_Inference Directory

The code under this directory is implementation of 17 S&P paper "Membership Inference Attacks Against Machine Learning Models" by Shokri et al.

• data_synthesis.py

It aims to implement Algorithm 1: Data Synthesis in paper.

• shadow_models.py

It aims to implement Shadow model technique in paper.

• attack_models.py

It aims to implement Membership Inference Attack according to paper.

Limitation

• I have only implemented Membership Inference Attack with classifier RandomForest in Python3 package sklearn. In the future, I will add various classifiers, like Neural Network and so on.

• In my code, I just implement the experiments. However, the code should be wrapped in a class for reusing.

Reference

1. Paper

• [17 S&P] Membership Inference Attacks Against Machine Learning Models

• [15 IJSN Attribute Infenrence Attack] Hacking Smart Machines with Smarter Ones How to Extract Meaningful Data from Machine Learning Classifiers

2. Code

Membership Inference Attack:

• Paper code: https://github.com/csong27/membership-inference

• BielStela's experiment implementation: https://github.com/BielStela/membership_inference