MSF-ADV

MSF-ADV is a novel physical-world adversarial attack method, which can fool the Multi Sensor Fusion (MSF) based autonomous driving (AD) perception in the victim autonomous vehicle (AV) to fail in detecting a front obstacle and thus crash into it.

Parper

Invisible for both Camera and LiDAR: Security of Multi-Sensor Fusion based Perception in Autonomous Driving Under Physical-World Attacks

Author: Yulong Cao*, Ningfei Wang*, Chaowei Xiao*, Dawei Yang*, Jin Fang, Ruigang Yang, Qi Alfred Chen, Mingyan Liu, Bo Li (*Co-first authors)

Website: https://sites.google.com/view/cav-sec/msf-adv

This is the code for the paper Invisible for both Camera and LiDAR: Security of Multi-Sensor Fusion based Perception in Autonomous Driving Under Physical-World Attacks accepted by IEEE S&P 2021.

The arxiv link to the paper: https://arxiv.org/abs/2106.09249

Installation

Install the required environments with the requirements.txt file using ANACONDA

$ conda env create -f requirements.yml

Command line

Download the target model

You can find the model through the official Baidu Apollo GitHub and YOLO website. Or you can download the model here and then unzip it to ./data folder.

Generating the adversarial object

python attack.py [-obj] [-obj_save] [-lidar] [-cam] [-cali] [-e] [-o] [-it]

Argument	Description
`-e`	Constrained max changing for the object vetex
`-o`	Optimization method: pgd and adam
`-it`	Max iteration number
`-obj`	Initial benign 3D object path
`-obj_save`	Adversarial 3D object saving dir
`-lidar`	LiDAR point cloud data path
`-cam`	Camera image data path
`-cali`	Calibration file path

Example for generating the adversarial object

python attack.py -obj ./object/object.ply -obj_save ./object/obj_save -lidar ./data/lidar.bin -cam ./data/cam.png -cali ./data/cali.txt -e 0.2 -o pgd -it 1000

Evaluation

The source code for evaluating the generated adversarial 3D object are in evaluation folder.

Citation

If you use this code for your research, please cite our papers.

@inproceedings{sp:2021:ningfei:msf-adv,
  title={{Invisible for both Camera and LiDAR: Security of Multi-Sensor Fusion based Perception in Autonomous Driving Under Physical World Attacks}},
  author={Yulong Cao and Ningfei Wang and Chaowei Xiao and Dawei Yang and Jin Fang and Ruigang Yang and Qi Alfred Chen and Mingyan Liu and Bo Li},
  booktitle={Proceedings of the 42nd IEEE Symposium on Security and Privacy (IEEE S\&P 2021)},
  year={2021},
  month = {May}
}

@inproceedings{xiao2019meshadv,
  title={{MeshAdv: Adversarial Meshes for Visual Recognition}},
  author={Xiao, Chaowei and Yang, Dawei and Li, Bo and Deng, Jia and Liu, Mingyan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6898--6907},
  year={2019}
}

Acknowledgments

Our code is inspired by Towards Robust LiDAR-based Perception in Autonomous Driving: General Black-box Adversarial Sensor Attack and Countermeasures, Mitsuba Physically Based Renderer, neural_renderer, pytorch-caffe, and PerceptualSimilarity.

ASGuard-UCI/MSF-ADV