yahiaelgamal/adversarial_noise

This package adds adversarial noise to image to trick models into predicting a target class

Image for WIP script that activates the output of a specific conolution (the specific image is from resnet152, layer_name='pretrained_model.layer2.0.conv1', activation_index_tuple=(0, 40, ))

---

Setup

0. cd to this repo
1. (optional) use pyenv/conda/virtualenv
2. pip install -r requirements.txt
3. pip install -e .

python adversarial_noise/model.py --image_path 'input_images/example_image4.jpg' \
  --target_class volcano \
  --output_image_path output_image.png \
  --model_name resnet152 \
  --output_intermediary_images  True\
  --output_intermediary_noise True

Example Usage

python adversarial_noise/model.py --help

Usage: model.py [OPTIONS]

Options:
  --image_path PATH               [required]
  --target_class TEXT             [required]
  --output_image_path PATH        [required]
  --model_name TEXT
  --max_iterations INTEGER
  --output_intermediary_images BOOLEAN
  --output_intermediary_noise BOOLEAN
  --help                          Show this message and exit.

python adversarial_noise/model.py --image_path 'input_images/example_image.jpg' --target_class volcano --model_name resnet152 python adversarial_noise/model.py --image_path 'input_images/example_image.jpg' --target_class volcano --model_name resnet152

The package supports imagenet classes and image transformation (mainly resizing, cropping and normalization). A list of classes in file imagenet_classes.txt

Two parameters that are would control how fast the process is, num_iteration (which can be set in the the cli interface), and a hardcoded parameter EARLY_STOP_LOSS_DELTA in adversarial_noise/model.py

Example output

More data in example_output folder in this repo.

python adversarial_noise/model.py --image_path 'input_images/example_image4.jpg' \
  --target_class volcano \
  --output_image_path output_image.png \
  --model_name resnet152 \
  --output_intermediary_images False \
  --output_intermediary_noise False