BTS: Consensus-Driven Anomaly Detection for Multi-Drone Perception

This repository contains the official implementation of the BTS (Build Trust in Swarm) framework and the CoSwarm dataset proposed in the paper "Building Trust in the Swarm: Consensus-Driven Anomaly Detection for Multi-Drone Perception".

Abstract

Multi-drone collaborative perception is important for secure and efficient low-altitude applications. Trust among drones is a foundation for reliable environmental awareness. However, shared perceptual data is vulnerable to adversarial attacks, which can damage integrity and break collaboration. Existing defenses are often not developed for dynamic drone swarms. They are costly, depend on heuristics, or lack a clear way to reach consensus and build trust. To address these gaps, we propose \underline{B}uild \underline{T}rust in \underline{S}warm (BTS), a consensus-driven anomaly detection framework for drone swarms. BTS integrates a distributed anomaly detection algorithm, enabling each drone to identify suspicious attackers through the consistency of perception results, combined with the Practical Byzantine Fault Tolerance (PBFT) mechanism to make consensus on excluding malicious nodes. We also introduce CoSwarm, a large-scale multi-modal dataset capturing bird’s eye view (BEV) perception from drones and roadside units (RSUs) in complex traffic scenarios for security-oriented collaborative perception. Extensive simulations show that BTS achieves over 84% latency reduction and improves the average precision of collaborative perception by 9.7% over state-of-the-art methods, and enhances the security and robustness of multi-drone systems in critical tasks.

Dataset: CoSwarm

Key Features

• Scale: 10,000 frames (100 scenes × 100 frames/scene) of synchronized LiDAR/RGB data.
• Sensors: 32-channel LiDAR (100m range, 1.2M points/s) and RGB cameras (800×800 resolution, 110° FoV) on drones (40m altitude) and RSU (15m altitude).
• Annotations: 3.18M 3D bounding boxes, 23 object classes, 8 attribute types, and drone trajectories.
• Scenarios: Dynamic urban traffic (200+ vehicles) in CARLA’s Town03/Town05, supporting 4-drone or 7-drone swarms.

Download

The CoSwarm dataset is available on Hugging Face. Download via the Hugging Face Dataset Repository

Extraction

Follow these steps to verify and extract the dataset:

1. Verify file integrity with md5sum -c checksums.md5.
2. Extract large split archives (e.g., samples/sweeps):

   cat samples.tar.gz.* | pigz -d | tar -xvf -
   cat sweeps.tar.gz.* | pigz -d | tar -xvf -

3. Extract small archives:

   tar -xzvf v1.0.tar.gz
   tar -xzvf maps.tar.gz

Code Implementation

1. Environment Setup and Installation

Clone this repository BTS:

git clone https://github.com/your-username/BTS.git
cd BTS

And then follow the Installation doc

2. Usage

2.1 Training

Train the BTS framework on the Parsed CoSwarm dataset (supports DDP or DP with multiple GPUs):

If you don't want use the default arguments, you can enter the training file to modify the default parameters of the code or add command line parameters

• DP: train_bts.py

python train_bts.py

• DDP: train_bts_ddp.py (default 4 GPUs)

python train_bts_ddp.py

2.2 Test

Run anomaly detection and secure collaborative perception on test data:

# coperception/tools/det/BTS
python BTS.py

Parameter settings

I. Basic Mandatory Parameters (Confirm Every Run)

Parameter	Description	Example Value
--data/-d	Path to dataset (replace {your location})	-d /{your location}/CoSwarm-det
--num_agent	Total number of collaborative agents (4~7 Drones + 1 RSU)	--num_agent 6
--com	Feature fusion strategy	--com max

II. Training-Related Parameters

Parameter	Description	Example Value
--nepoch	Total training epochs	--nepoch 50
--batch	Scene-level batch size	--batch 4
--lr	Initial learning rate	--lr 0.001
--resume	Path to saved model for resuming training	--resume /{your location}/epoch_50.pth
--log	Enable logging (no value needed; add the flag to activate)	--log
--logpath	Path to save log files	--logpath /{your location}/logs

III. Adversarial Attack & Defense Parameters

Parameter	Description	Example Value
--adv_method	Adversarial attack method	--adv_method pgd
--eps	Perturbation strength	--eps 1.0
--pbft	PBFT defense mode	--pbft pbft_mAP
--tolerant	Number of tolerable malicious nodes	--tolerant 1

IV. Inference & Visualization Parameters

Parameter	Description	Example Value
--scene_id	Target evaluation scene ID	--scene_id 81
--visualization	Visualize inference results (add the flag to activate)	--visualization
--ego_agent	ID of the ego agent	--ego_agent 3

Results

Main Performance

Comparison of defense performance between BTS and different baselines

Baseline	AP@IoU=0.5	AP@IoU=0.7
IUB	73.7	66.6
UB (oracle)	71.2	64.8
BTS (ours)	71.1	65.0
ROBOSAC	65.3	59.3
LB	55.7	50.0
ND	23.1	22.2

Paper Citation

If you use BTS or the CoSwarm dataset in your research, please cite our paper:

need to add

Acknowledgment

BTS is modified from coperception library.

The collection script of CoSwarm is modified from carla_nuscenes

This project is not possible without these great codebases.

Contact

For questions, please contact in issues.