FGSM Brute Force Attack

This project demonstrates a Fast Gradient Sign Method (FGSM) brute force attack on image classification models hosted on Hugging Face 🤖. It generates adversarial examples by perturbing input images, attempting to force the model into misclassifying them.

Supports targeted attacks by specifying a desired misclassification label, and allows testing across different epsilon values.

Installation

pip install -r requirements.txt

Usage

The script takes the following arguments:

  --model: The name of the Hugging Face model.
  --image: The path to the input image.
  --epsilon: The epsilon value for the attack (default: 0.05).
  --size: The size of the input image for the model.
  --target: The target class name.
  --d: The directory to save adversarial images.