Countering Adversarial Images using Input Transformations

Paper: Guo et al. 2018

Setup

Run ./setup.sh to fetch models.

Breaks

• Bit-depth reduction: bitdepth.ipynb (broken with BPDA)
• JPEG: jpeg.ipynb (broken with BPDA)
• Cropping: crop.ipynb (broken with EOT)
• Quilting: quilt.ipynb (broken with EOT+BPDA)
• Total variation denoising: tv.ipynb (broken with EOT+BPDA)

robustml evaluation

Run with:

python robustml_attack.py --imagenet-path <path> --defense <defense>

Where <defense> is one of bitdepth, jpeg, crop, quilt, or tv.