/robust_physical_perturbations

Public release of code for Robust Physical-World Attacks on Deep Learning Visual Classification (Eykholt et al., CVPR 2018)

Primary LanguagePythonMIT LicenseMIT

This repository holds the code (and some results) used in Robust Physical-World Attacks on Deep Learning Visual Classification. The software carries an MIT license.

The folders are as follows:

  • lisa-cnn-attack holds the code to attack the LISA-CNN that classifies US road signs from the LISA dataset. Contains a model that achieves 91% accuracy on that dataset. This is the most rudimentary implementation of the algorithm.
  • gtsrb-cnn-attack holds the code that attacks the GTSRB-CNN that classifies German road signs (with the stops replaced with US ones from LISA). Implementation somewhat improved.
  • imagenet-attack holds the code that attacks the Inception V3 model that operates on ImageNet data. Most advanced implementation of the algorithm.

Further details are given in README files in the respective folders. They also specify how to download portions that are needed for the code to run but are not committed here due to size.

Note that in lisa-cnn-attack and in gtsrb-cnn-attack we include portions of an older version of the cleverhans library for compatibility. It carries its own MIT License.