This repository is for NeurIPS 2018 spotlight paper Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples.

Update: for Pytorch implementation, please refer to src/pytorch.

• opencv-python
• dlib
• caffe

• Please download VGG-Face caffe model from here.
• Unzip the model under data/ folder.

In attribute_mutation.ipynb, attribute-substituted and attribute-preserved images are produced for the base image. Four attributes are encoded with indices from 0 to 3. See the following table for details. Please use attributes[index] for corresponding attributes.

Two actions are also encoded with indices, which is listed in the following table. Please use actions[index] for corresponding actions.

Generated images are saved in folder data/attribute_mutated/[attribute]_[action]/.

Attribute witnesses are extracted layer by layer based on attribute-substituted and attribute-preserved images. Please find the implementation in witness_extraction.ipynb. Extracted witnesses are saved in folder data/witnesses/.

With extracted attribute witnesses, neuron weakening and strengthening are applied for each input during execution. Adversary detection is achieved by observing the final prediction from attribute-steered model comparing to the original model. Detailed implementation is in adversary_detection.ipynb.

7 adversarial attacks are included in folder data/attacks. Please change attack_path in the code to test on different attacks.

Please cite for any purpose of usage.

@inproceedings{NeurIPS2018_7998,
    title={Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples},
    author={Tao, Guanhong and Ma, Shiqing and Liu, Yingqi and Zhang, Xiangyu},
    booktitle={Advances in Neural Information Processing Systems 31},
    pages = {7728--7739},
    year={2018}
}