CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird's Eye View Perception

👥 Authors

Name	Profile
Senkang Hu*	Google Scholar
Yihang Tao*	Google Scholar
Guowen Xu	Google Scholar
Yiqing Deng	Google Scholar
Xianhao Chen	Google Scholar
Yuguang Fang	Google Scholar
Sam Kwong	Google Scholar

📝 Abstract

Collaborative Perception (CP) has shown a promising technique for autonomous driving, where multiple connected and autonomous vehicles (CAVs) share their perception information to enhance the overall perception performance and expand the perception range. However, in CP, ego CAV needs to receive messages from the collaborators, which makes it easy to be attacked by malicious agents. For example, a malicious agent can send harmful information to the ego CAV to mislead it. To address this critical issue, we propose a novel method, CP-Guard, a tailored defense mechanism for CP that can be deployed by each agent to accurately detect and eliminate malicious agents in its collaboration network. Our key idea is that CP will lead to a consensus rather than a conflict against the ego CAV's perception results. Based on this idea:

• We develop a probability-agnostic sample consensus (PASAC) method that can effectively sample a subset of the collaborators and verify the consensus without prior probabilities of malicious agents
• We design a collaborative consistency loss (CCLoss) to calculate the discrepancy between the ego CAV and the collaborators
• We conduct extensive experiments in collaborative bird's eye view (BEV) tasks

🛠️ Installation

System Requirements

Requirement	Version/Specification
OS	Linux (tested on Ubuntu 22.04)
Python	3.7
Package Manager	Miniconda
Deep Learning Framework	PyTorch
CUDA	11.7

Setup Steps

1. Create and Activate Conda Environment

cd coperception
conda env create -f env.yml
conda activate coperception

2. Install CUDA Dependencies

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

3. Install CoPerception Library

pip install -e .

📊 Dataset Preparation

1. Download and unzip the parsed segmentation dataset of V2X-Sim 2.0

2. Link the test split of V2X-Sim dataset in the default value of argument "data":

/{Your_location}/V2X-Sim-seg/test

Dataset Structure

test
├──agent_0
├──agent_1
├──agent_2
├──agent_3
├──agent_4
├──agent_5
      ├──5_0
      ├──0.npy
      ...

Model Checkpoint

1. Download pre-trained weights
2. Save them in CP-Guard/coperception/ckpt/meanfusion folder

🚀 Running CP-Guard

Basic Usage

cd coperception/tools/seg/
python cp_guard.py [options]

Note: Due to data syncing issues, CP-Guard cannot be performed under multi-GPU environment. For multiple GPUs, use: CUDA_VISIBLE_DEVICES=0 python cp_guard.py [options]

Parameters

General Parameters

Parameter	Description	Default
-d, --data	Path to preprocessed sparse BEV training data	./datasets/V2X-Sim-seg/test
--batch	Number of scene	1
--th	CCLoss threshold	0.08
--com	Communication method (disco/when2com/v2v/sum/mean/max/cat/agent)	v2v
--num_agent	Total number of agents	6

Adversarial Attack Parameters

Parameter	Description	Default
--pert_alpha	Scale of the perturbation	0.1
--adv_method	Attack method (pgd/bim/cw-l2/fgsm)	pgd
--eps	Epsilon of adv attack	0.5
--adv_iter	Adv iterations of computing perturbation	15

Scene Settings

Parameter	Description	Default
--scene_id	Target evaluation scene (Scene 8, 96, 97 has 6 agents)	[8]
--sample_id	Target evaluation sample	None

CP-Guard Specific Parameters

Parameter	Description	Default
--cpguard	Mode (upperbound/lowerbound/no_defense/pasac/robosac)	no_defense
--ego_agent	ID of ego agent	1
--ego_loss_only	Only use ego loss to compute adv perturbation	False
--number_of_attackers	Number of malicious attackers	1
--fix_attackers	Fix attackers across frames	False

ROBOSAC Specific Parameters

Parameter	Description	Default
--robosac_k	Consensus set size	None
--step_budget	Sampling budget per frame	5
--use_history_frame	Use history frame for consensus	False

🙏 Acknowledgment

CP-Guard is modified from:

• coperception library
• robosac library

📚 Citation

If you find this project useful in your research, please cite:

@article{CP_Guard_2025, 
    title={CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird's Eye View Perception}, 
    volume={39},
    number={22},
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
    author={Hu, Senkang and Tao, Yihang and Xu, Guowen and Deng, Yiqin and Chen, Xianhao and Fang, Yuguang and Kwong, Sam}, 
    year={2025},
    month={Apr.},
    pages={23203-23211} 
}