Poisoned ML Model Demo

Demonstrates the training and exploitation of a poisoned machine learning model.

Introduction

This repository contains scripts for generating adversarial patches, injecting these patches into image datasets, and displaying a video feed overlaid with the output of an object classification algorithm. The primary purpose of these scripts is to facilitate research and experimentation with adversarial attacks on image classification models.

Project Overview

The project comprises three main scripts:

1. generate_patch.py: Generates an adversarial patch and saves it as an image file.
2. poison_data_class.py: Injects an adversarial patch into every image in a specified directory.
3. feed.py: Displays a video feed with the output of an object classification model overlaid on the video.

File Descriptions

generate_patch.py

This script generates an adversarial patch by randomly filling a grid with black rectangles and saves the result as a PNG image.

Usage:

python3 generate_patch.py <scale_factor> <output_file>

Dependencies:

• pillow

poison_data_class.py

This script injects an adversarial patch into every image in a specified input directory and saves the modified images to an output directory. The patch is resized and added at a random position in each image, with Gaussian noise applied to enhance robustness.

Usage:

python3 poison_data_class.py <adversarial_patch_path> <input_dir> <output_dir>

Dependencies:

• pillow
• numpy
• tqdm

feed.py

This script captures a video feed, processes each frame with an object classification model, and overlays the classification result on the video feed. The model is loaded from a specified file, and the video feed is captured from a specified device.

Usage:

python3 feed.py [model_file=model.keras] [device_id=0] [tolerance=0.99]

Dependencies:

• tensorflow/keras
• OpenCV
• numpy
• matplotlib

Setup Instructions

Prerequisites

Ensure you have the following installed:

• Python 3.x
• pip (Python package installer)

Installing Dependencies

You can install the required Python packages using pip:

pip install -r requirements.txt

Running the Scripts

1. Generate an Adversarial Patch

   To generate an adversarial patch, run:

   python3 generate_patch.py <scale_factor> <output_file>

   Replace <scale_factor> with the desired scale factor for the patch and <output_file> with the path to save the generated patch.

2. Inject the Adversarial Patch into a Dataset

   To inject the generated patch into a dataset, run:

   python3 poison_data_class.py <adversarial_patch_path> <input_dir> <output_dir>

   Replace <adversarial_patch_path> with the path to the generated patch, <input_dir> with the path to the input image directory, and <output_dir> with the path to save the modified images.

3. Display Video Feed with Classification Overlay

   To display the video feed with the classification overlay, run:

   python3 feed.py [model_file=model.keras] [device_id=0] [tolerance=0.99]

   Optionally, replace [model_file] with the path to your model file, [device_id] with the ID of your camera device, and [tolerance] with the classification tolerance threshold.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Contact

Author: Saul Johnson
Email: saul.johnson@nhlstenden.com

For any questions or issues, please open an issue on this repository.