Adversarial Noise

This repository contains code that adds adversarial noise to a given image, tricking an image classification model into misclassifying the altered image as a specified target class.

🔨 Requirements and Installation

Miniforge (Conda) is recommended for creating a Python environment.

# Clone the repo
git clone https://github.com/rsundar96/adversarial-attack.git

# Create a Python environment with all of the required dependencies
conda env create -f environment.yml -v

# (For local development only)
# Setup pre-commit for code formatting
pre-commit install

# Activate the environment
conda activate adversarial-attack

🧑‍💻 Local Development

Black formatting should be applied on all code before making a PR - if properly setup, this is covered by pre-commit.

When updating dependencies, the requirements.txt file also needs to be updated in addition to pyproject.toml. The following command should update the requirements.txt file.

pip-compile --extra=dev --output-file=requirements.txt pyproject.toml

🧪 Tests

Test framework in use is pytest. To check code coverage of tests, run the following commands

# Run tests using code coverage
coverage run --module pytest tests

# Generate test report
coverage report -m -i
coverage html -i

🚀 Usage

A few sample images are present under the images folder. You may use these images, or upload images into this folder.

The list of target classes currently supported are those present in the ImageNet dataset.

# Example to make the pre-trained ResNet-18 image classification model misclassify a panda as a gibbon
python src/add_adversarial_noise.py --image images/panda.jpeg --target-class "gibbon"