jxcwu / magnet

MagNet: a Two-Pronged Defense against Adversarial Examples

MagNet

Demo code for "MagNet: a Two-Pronged Defense against Adversarial Examples", by Dongyu Meng and Hao Chen, at CCS 2017.

The code demos black-box defense against Carlini's L2 attack of various confidences. Other techniques proposed in the paper are also included in defensive_models.py and worker.py, but are not shown in the demo defense. Attack implementations are not provided in this repository.

Run the demo code:

0. Make sure you have Keras, Tensorflow, numpy, scipy, and matplotlib installed.
1. Clone the repository.
2. We provide demo attack data and classifier on Dropbox and 百度网盘 (密码: yzt4). Please download and put the unzipped files in MagNet/. You may also use your own data for test.
3. Train autoencoders with python3 train_defense.py.
4. Test the defense with python3 test_defense.py .
5. Defense performance is plotted in graph/defense_performance.pdf.