black-box-attacks

Comparison of gradient estimation techniques for black-box adversarial examples

This is a fork of https://github.com/labsix/limited-blackbox-attacks

Link to results: http://www.homepages.ucl.ac.uk/~ucabaye/papers/black_box_attacks.pdf

Choices of gradient estimation are: NES, RDSA, SPSA, SPSA (1-sided)

To run:

1. Download Make a directory tools/data, and in it put the decompressed Inceptionv3
   classifier from (http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz)

2. Set IMAGENET_PATH and METADATA_PATH in main.py and attacks.py to the location of the ImageNet dataset on your machine.

3. To run all experiments: ./query.sh