ML-LOO: Detecting Adversarial Examples with Feature Attribution

Code for ML-LOO on a minimal example.

Dependencies

The code runs with Python 2.7 and requires Tensorflow 1.11.0, and Keras 2.2.4. Please pip install the following packages:

• numpy
• scipy
• tensorflow
• keras
• scikit-learn

Preparation stage

Generate the adversarial examples by C&W attack, to be used in the training stage of ML-LOO

###############################################
python generate_attack.py --dataset_name cifar10 --model_name resnet
###############################################

Extract ML-LOO features (i.e., IQR of LOO attribution maps of different layers of a neural network), and then split the data set containing original and adversarial examples into a training set and a test set.

###############################################
python features_extract.py --dataset_name cifar10 --model_name resnet --attack cw --det ml_loo
###############################################

Train ML-LOO and evaluate its performance

Train ML-LOO and evaluate it on the test set.

python train_and_evaluate.py --dataset_name cifar10 --model_name resnet --data_sample x_val200

The generated AUC plot can be found in cifar10resnet/figs/.