Adversarial Example Detection using Maximum Mean Discrepancy (MMD) and Bootstrap Hypothesis Test.

The adversarial dataset and the classification model was created by another our project in this link: Adversarial Robustness of Medical Image Classifiers via Denoised Smoothing

Overview

This project aims to detect adversarial examples in datasets using the Maximum Mean Discrepancy (MMD) and Bootstrap Hypothesis test.

Adversarial examples are inputs to a machine learning model that are designed to cause the model to misbehave. The MMD test is a statistical test that measures the difference between two distributions.

Results

Code Structure

• hypothesis_test.py: This file contains the implementation of the MMD test and related functions.
• create_dataset.py: This file contains functions for creating and manipulating datasets.
• data_utils.py: This file contains utility functions for working with data.
• test.py: This file contains the main script for running the MMD test on a dataset.

Functionality

• The project provides a function test_distribution_difference that takes in two datasets and returns the MMD value, p-value, and a boolean indicating whether the null hypothesis (i.e., the two distributions are the same) is rejected.
• The project also provides functions for creating and manipulating datasets, including create_dataset and mix_datasets.
• The main function in test.py runs the MMD test on a dataset and prints the results.

Usage

• Run the MMD test:

  python test.py --adversarial_attack <attack_type> --adversarial_percentage <percentage> --test_size <size> --n_interations <iterations> --test <test_number>

• --adversarial_attack: type of transforms (FGSM, PGD, flips, subsampling, gaussian_blur)
• --adversarial_percentage: adversarial percentage of full dataset(0.0 - 1.0)
• --test_size: Size of Y dataset
• --n_interations: Number of bootstrap iterations

Requirements

• Python 3.x
• PyTorch
• NumPy
• SciPy
• Argparse

Notes

• The project assumes that the datasets are stored in a specific format (e.g., PyTorch tensors).
• The project uses the argparse library to parse command-line arguments.