Adversarial Neural Network Toolkit

The source code in the repository is a Python tool that is able to fulfil a complete workflow of:

1. Traning full-precision and quantized neural networks.
2. Creating adversarial examples on these networks.
3. Transfering adversarial examples from the network where samples are created (source) to another (target) network.

The tool can be used in two ways:

1. As a standalone Python based console application which executes based on the configurations specified in a YAML file (ANNT/config.yaml). More details: wiki.
2. As a Python library that can be imported. Individual modules (training, inference, creating adversarial examples, and transfer of adversarial examples) can be then done within a Python script.
   For this, please follow the examples in the .ipynb files in the root.

---

The API uses Tensropack (Y. Wu et al., 2016) for training. Tensorpack is training framework which is a part of TensorFlow 1.13 (Abadi et al., 2016) API.
For quantization DoReFa-Net method (Zhou et al., 2018) is used.
For adversarial attack generation ART (Nicolae et al., 2019) is used.

---

References

Quantization is based on DoReFa-Net method as proposed in the paper: https://arxiv.org/abs/1606.06160

Cited as:
Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., & Zou, Y. (2018). DoReFa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv:1606.06160 [cs].

For quantization, the library available from the authors is used. This is available at:

https://github.com/tensorpack/tensorpack/tree/master/examples/DoReFa-Net

The networks used are from the examples provided on the Tensorpack repository:
Tensorpack cited as:
Wu, Y. et al. (2016). Tensorpack. https://github.com/tensorpack.

Tensorpack models are available at:

https://github.com/tensorpack/tensorpack/tree/master/examples

Adversarial Examples are created using Adversarial Robustness Toolbox (ART) v. 1.5.1. Official paper: https://arxiv.org/abs/1807.01069

Cited as:
Nicolae, M.-I., Sinn, M., Tran, M. N., Buesser, B., Rawat, A., Wistuba, M., Zantedeschi, V., Baracaldo, N., Chen, B., Ludwig, H., Molloy, I. M., & Edwards, B. (2019). Adversarial robustness toolbox v1.0.0. arXiv:1807.01069

ART is one of the popular APIs for adversaral examples generation and supports a large number of attacks. It is open-source with large number of very well explained examples. Please check out their repository at: https://github.com/Trusted-AI/adversarial-robustness-toolbox