code-book-defenses

Evaluation of adversarial robustness of neural networks that use MSE and code-book target representations

Setting up

We use Python 3.x. Install the dependencies via:

$ pip3 install -r requirements.txt

Note that we depend on tensorflow-gpu==1.19.0. This may not work depending on your system's configurations and library/driver versions (e.g. CUDA, CuDNN). Please install the appropriate TensorFlow version.

Training & attacking a model

• The main.py module trains a DenseNet model and runs a number of untargeted and targeted attacks on it during test time.
• The JSON files in configs/ are the configurations of the DenseNet model. They specify parameters such as architecture of the model, dataset to train on, etc.
• code_book_defenses/config.py includes several variables and dictionaries which configure the attack experiments.
• For example, the lists untargeted_attacks and targeted_attacks indicate which attacks to run on the trained model.
• The attack_name_to_params dictionary enumerates the parameters for each attack. You will notice that at most one field from every attack dictionary is a list. For each attack during test time, we iterate over each value of the list to construct adversarial examples.
• N.B. The Carlini Wagner L2 and L-BFGS attacks take a very long time (> 24 hours) especially for all test data, which is 10,000 images. If you want to run quick experiments, you can set the number of images to a smaller number at the top of code_book_defenses/config.py.

After you have installed the dependencies and identified the configurations you want to run, you can trigger training via the following:

$ python3 main.py --config=configs/name_of_config.json --gpus=0 --experiment=experiment_1

The arguments for main.py are as follows:

• config (required): The path to the configuration JSON file
• gpus: GPUs to use (NOTE: the current implementation is not optimized for multi-gpu training)
• experiment: Name of the experiment (can be some unique identifier)