AdvCollaborativePerception

Repo for USENIX security 2024 paper "On Data Fabrication in Collaborative Vehicular Perception: Attacks and Countermeasures" https://arxiv.org/abs/2309.12955

Requirements

• Python package manager: conda
• GPU (tested on RTX 2080 Ti)
• Free space > 40 GB

Get started

git clone --recursive https://github.com/zqzqz/AdvCollaborativePerception.git
cd AdvCollaborativePerception
bash scripts/setup.sh
conda activate advCP
python scripts/evaluate.py
cat result/evaluate.log

The scripts will install dependencies, download dataset from Google Drives, set up the environment, and execute all evaluation tasks. Results are saved to result by default.

Structure of the repository

• data/OPV2V: The perception dataset OPV2V and pre-computed meta files.
• data/carla: Pre-computed carla maps where OPV2V is collected.
• data/model_3d: 3D models that are useful for ray casting attacks.
• models: Pretrained models.
• mvp: The main module implementing our proposed attack and defense.
• test: Examples of using mvp to operate datasets, attack methods, etc.

Debugging information

Q: Problems of CUDA and PyTorch.

A: Our script scripts/setup.sh by default installs PyTorch 1.9.1 and CUDA 10.2. Please adjust the versions if it does not work for your environment.

Q: Deprecated functions of numpy or shapely.

A: Our code is tested on numpy==1.19 and shapely==1.8.1. numpy>=1.20 and shapely>=2.0 may throw warnings or errors about deprecated functions.