PercepGuard

This repository demonstrates the LSTM bounding box sequence classifier of our PercepGuard paper:

@inproceedings{man2023percepguard,
  title={That Person Moves Like A Car: Misclassification Attack Detection for
         Autonomous Systems Using Spatiotemporal Consistency},
  author={Man, Yanmao and Muller, Raymond and  Li, Ming and Celik, Z. Berkay
          and Gerdes, Ryan},
  booktitle={USENIX Security Symposium},
  year={2023}
}

The scripts are written in Python 3, with dependencies

tensorflow == 2.1.0
numpy == 1.18.5

The pre-trained model for the BDD100K MOT dataset can be downloaded from Google Drive. Unzip it into models:

mkdir models
unzip /path/to/bdd100k.zip -d models

To use the pre-trained model, see demo.py for a simple example:

python3 demo.py

This outputs [[0.01 0.04 0.82 0.01 0.12]], where the car category has the highest score.