CSED415: Computer Security

Professor: Seulbae Kim
Team: whysw (Chiheon Kim, Sungjae Cho)

Adversarial Attacks on Traffic Sign Recognition in a Physical Domain

Note: This repository is forked from the original author's repository.

Reproduce and further explore ShadowAttack(paper / github)

New features

• Seed Fix
  Add seed_everything() for reliable reproductions. A seed can be specified in params.json.

• Video Preprocess
  video_preprocess.py saves cropped traffic sign images from each frame of given video. (The video should be located under videos/ directory and have .mp4 format.) It requires a JSON file exported from "Lable Studio", where you can manually label a traffic sign with a rectangle box. With the key frame information in the JSON file, it calculates interpolation and finds the position of the traffic sign in each frame. Then it saves the images under videos/<video_file_name>-frames/<frame#>.jpg.

  python video_preprocess.py test.mp4

• Automated single_image_test() for given directory
  To test every single image saved under videos/<video_file_name>-frames/, lisa.py is slightly modified. Given a directory name under videos/, it iteratively test each image and saves all the results in specified log file.

  python lisa.py test-frames test.log