ODPure: Backdoor Purification for Object Detection via Ensemble Corruption Consensus

The first input-stage black-box purification defense tailored for object detectors against backdoor attacks.

📖 Overview

ODPure is the first input-stage, black-box purification framework tailored for defending object detectors against backdoor attacks. It operationalizes a novel Corruption-Reconstruction-Selection (CRS) paradigm that:

Corrupts input images with diverse perturbations to disrupt trigger patterns
Reconstructs fine-grained structural features using generative diffusion priors
Selects high-confidence detections via spatial consensus voting

ODPure effectively neutralizes diverse backdoor attacks (reducing Attack Success Rate to as low as 0.0%) while preserving clean detection utility, achieving superior defense-utility trade-offs over existing defenses.

🔑 Key Features

Black-Box Defense: No need for model internals, gradients, or architectures
Model-Agnostic: Works with any object detector (YOLO, Faster R-CNN, etc.)
Multi-Attack Defense: Handles OMA, ODA, and OGA attack types
Zero-Shot: No prior knowledge of trigger patterns required
Continuous Perception: Maintains operational stream continuity without discarding data

📋 Requirements

System Requirements

Python 3.8+
CUDA 11.7+ (for GPU acceleration)
16GB+ GPU memory recommended

Python Dependencies

# Core dependencies
torch>=1.13.0
torchvision>=0.14.0
numpy>=1.21.0
Pillow>=9.0.0
opencv-python>=4.5.0

# Diffusion models
diffusers>=0.14.0
transformers>=4.25.0
accelerate>=0.20.0

# Evaluation
scikit-learn>=1.2.0  # For DBSCAN clustering
scipy>=1.9.0

# Data processing
pyyaml>=6.0
tqdm>=4.64.0

Conda Environment Setup (Recommended)

# Create conda environment
conda create -n odpure python=3.9 -y
conda activate odpure

# Install PyTorch with CUDA
conda install pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia

# Install other dependencies
pip install -r requirements.txt

📦 Download Pretrained Weights

Before running the Reconstruction module, you need to download the pretrained model weights:

Required Weights

Installation Steps

Create the weights directory:

mkdir -p Method/Reconstruction/weights

Download and place the weights following the instructions in the respective model repositories.
Update the config files in Method/Reconstruction/configs/ to point to your downloaded weights.

🏗️ Project Structure

ODPure/
├── attack_script/              # Backdoor attack implementations
│   ├── COCO_chessboard_29x29_OMA.py     # Object Misclassification (Chessboard)
│   ├── COCO_chessboard_29x29_ODA.py     # Object Disappearance Attack (Chessboard)
│   ├── COCO_chessboard_9x9_OGA.py      # Object Generation Attack (Chessboard)
│   ├── COCO_poke_15x15_OMA.py          # OMA (Poké Ball)
│   ├── COCO_poke_15x15_ODA.py          # ODA (Poké Ball)
│   ├── COCO_poke_15x15_OGA.py          # OGA (Poké Ball)
│   ├── COCO_white_15x15_OMA.py         # OMA (Solid White)
│   ├── COCO_white_15x15_ODA.py         # ODA (Solid White)
│   └── COCO_white_15x15_OGA.py         # OGA (Solid White)
│
├── Method/                     # Core defense methodology
│   ├── corruptions/           # Image corruption module
│   │   ├── imagecorruption.py # Corruption functions
│   │   └── multiprocess_imagecorruption.py  # Parallel processing
│   │
│   ├── Reconstruction/         # Diffusion-based restoration
│   │   ├── inference.py       # Main inference script
│   │   ├── diffbir/           # DiffBIR model implementation
│   │   ├── llava/             # LLaVA captioner
│   │   ├── ram/               # Recognition-Aware Model
│   │   ├── configs/           # Model configurations
│   │   └── weights/           # Model weights (download separately)
│   │
│   └── dbscan_vote_new.py     # DBSCAN clustering & voting
│
├── evaluation/                 # Evaluation metrics
│   ├── OMA_ASR_new.py         # OMA Attack Success Rate
│   ├── OMA_mAP.py             # OMA Mean Average Precision
│   ├── ODA_ASR_new.py         # ODA Attack Success Rate
│   ├── ODA_mAP.py             # ODA Mean Average Precision
│   ├── OGA_ASR_new.py         # OGA Attack Success Rate
│   ├── OGA_mAP.py             # OGA Mean Average Precision
│   ├── val_OMA.py             # YOLO validation for OMA
│   ├── val_ODA.py             # YOLO validation for ODA
│   └── val_OGA.py             # YOLO validation for OGA
│
├── data_format_conversion/    # Data format utilities
│   ├── voc2yolo.py            # VOC to YOLO format conversion
│   ├── cocotoyolo.py          # COCO to YOLO format conversion
│   └── select_coco_val_attack_information.py
│
├── ablation_study/            # Ablation experiments
│   ├── multiprocess_imagecorruption.py
│   ├── random_select_corruption.py
│   └── select_specific_corruption.py
│
├── run_pipeline.sh            # One-shot CRS pipeline runner
└── README.md                 # This file

🚀 Quick Start

Option A: One-Shot End-to-End Pipeline (Recommended)

The recommended entry point is run_pipeline.sh, which cascades the three CRS stages:

# Defaults: GPU=0, INPUT=inputs/demo/bid, OUTPUT=results/v2.1_demo_bid, ATTACK=ODA
bash run_pipeline.sh

# Custom arguments: GPU_ID INPUT_DIR OUTPUT_DIR ATTACK
bash run_pipeline.sh 0 inputs/coco_oda results/oda ODA
bash run_pipeline.sh 1 inputs/coco_oma results/oma OMA
bash run_pipeline.sh 2 inputs/coco_oga results/oga OGA

The script performs:

Stage 1 (Corruption) – generates 45 corrupted variants
Stage 2 (Reconstruction) – DiffBIR v2.1 with LLaVA captioner
Stage 3 (Selection) – DBSCAN spatial consensus voting

Final purified detections are written to <OUTPUT_DIR>/final/.

Option B: Run Each Stage Manually

Use this if you want to inspect / replace intermediate steps.

Stage 1: Corruption

python Method/corruptions/multiprocess_imagecorruption.py \
    --input_dir /path/to/input/images \
    --output_dir /path/to/corrupted/images \
    --num_corruptions 45 \
    --num_workers 8

Stage 2: Reconstruction

python Method/Reconstruction/inference.py \
    --task denoise \
    --upscale 2 \
    --version v2.1 \
    --captioner llava \
    --cfg_scale 6 \
    --noise_aug 1 \
    --input /path/to/corrupted/images \
    --output /path/to/restored/images \
    --batch_size 32 \
    --device cuda

Stage 3: Selection (DBSCAN Voting)

python Method/dbscan_vote_new.py \
    --folder_purs /path/to/restored/detections \
    --temp /path/to/temp \
    --output_path /path/to/final/detections \
    --eps 0.5 \
    --min_samples 10

📊 Evaluation

Step 1: Activate Environment

conda activate odpure

Step 2: Run Detector on Three Splits

# Run on poisoned inputs (before defense)
python evaluation/val_ODA.py \
    --weights runs/train/exp/weights/last.pt \
    --data data/ODA_poison.yaml \
    --img 640 --iou-thres 0.65 --conf-thres 0.5 \
    --save-txt --save-conf \
    --project results/poisoned_val_txt

# Run on clean inputs
python evaluation/val_ODA.py \
    --weights runs/train/exp/weights/last.pt \
    --data data/ODA_clean.yaml \
    --img 640 --iou-thres 0.65 --conf-thres 0.5 \
    --save-txt --save-conf \
    --project results/clean_val_txt

# Run on purified inputs (after defense via ODPure)
python evaluation/val_ODA.py \
    --weights runs/train/exp/weights/last.pt \
    --data data/ODA_purified.yaml \
    --img 640 --iou-thres 0.65 --conf-thres 0.5 \
    --save-txt --save-conf \
    --project results/pur_val_txt

Step 3: Compute ASR / mAP

OMA (Object Misclassification Attack)

python evaluation/OMA_ASR_new.py

Configure paths inside the script:

gt_folder      = "/path/to/ground_truth"
benign_folder  = "/path/to/clean_val_txt"
attack_folder  = "/path/to/pur_val_txt"
target_class   = "0"      # person class

ODA (Object Disappearance Attack)

python evaluation/ODA_ASR_new.py

OGA (Object Generation Attack)

python evaluation/OGA_ASR_new.py

Compute mAP

python evaluation/ODA_mAP.py
python evaluation/OGA_mAP.py
python evaluation/OMA_mAP.py

🛡️ Defense Configuration

Default ODPure Settings

Corruption Types

ODPure uses 15 diverse corruption functions across 4 categories:

Noise: Gaussian, Shot, Impulse
Blur: Gaussian, Glass, Motion, Zoom, Defocus
Weather: Snow, Frost, Fog, Brightness
Digital: JPEG compression, Pixelate, Elastic transform

🔧 Advanced Usage

Custom Corruptions

from Method.corruptions.imagecorruption import *

# Apply specific corruption
corrupted_img = gaussian_noise(image, severity=2)
corrupted_img = glass_blur(image, severity=1)
corrupted_img = jpeg_compression(image, severity=3)

Multi-GPU Processing

# Process on specific GPU
CUDA_VISIBLE_DEVICES=0 python Method/Reconstruction/inference.py \
    --task denoise --input inputs/demo --output results/demo

# Batch processing with multiple GPUs
CUDA_VISIBLE_DEVICES=0,1,2,3 python Method/Reconstruction/inference.py \
    --task denoise --batch_size 64 --input inputs/batch --output results/batch

Integration with YOLOv5

# Train backdoored model
CUDA_VISIBLE_DEVICES="0,1" python train.py \
    --data data/custom.yaml \
    --epochs 200 \
    --weights checkpoints/yolov5s.pt \
    --img 640 \
    --batch-size 128

# Evaluate with defense
python evaluation/val_ODA.py \
    --weights runs/train/exp/weights/last.pt \
    --data data/val.yaml \
    --img 640 \
    --iou-thres 0.65 \
    --conf-thres 0.5 \
    --save-txt --save-conf

📈 Experimental Results

Defense Performance on VOC and COCO (Chessboard Trigger)

Robustness Against Adaptive Attacks

🔬 Threat Model

Attacker's Capability

Implants backdoor via data poisoning during training
Manipulates detectors to exhibit malicious behaviors (OMA, ODA, OGA)
Maintains normal predictions on clean inputs

Defender's Constraints

Attack Agnostic: No prior knowledge of trigger patterns
Strict Black-Box: Only query inputs and observe outputs
No Training Data: Purification on inference-time inputs only

Defense Goals

Security: Mitigate backdoor activation (low ASR)
Fidelity: Maintain high mAP on clean and purified inputs

Attack Types

Attack Type	Description	Behavior

📚 Citation

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

@article{odpure2026,
  title={ODPure: Backdoor Purification for Object Detection via Ensemble Corruption Consensus},
  author={},
  journal={},
  year={2026}
}

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Acknowledgements

DiffBIR - Diffusion-based Image Restoration
LLaVA - Large Language and Vision Assistant
YOLOv5 - Object Detection Framework
CLIP - Contrastive Language-Image Pretraining

📧 Contact

For questions or collaborations, please open an issue on GitHub.

ODPure - Protecting object detection systems from backdoor attacks while maintaining continuous perception.

DiffBIR v2.1	Main restoration model	HuggingFace
Stable Diffusion	Latent diffusion priors	HuggingFace
LLaVA	Vision-language captioner	HuggingFace
RAM	Recognition-Aware Model	GitHub Release

`num_corruptions`	15 types × 3 severities = 45 variants	Corruption diversity
`corruption_severity`	1, 2, 3	Corruption intensity levels
`DBSCAN eps`	0.5	Clustering radius
`DBSCAN min_samples`	10	Consensus threshold
`cfg_scale`	6.0	Classifier-free guidance
`noise_aug`	1	Noise augmentation level

OMA	VOC (YOLO)	76.4%	8.2% / 87.7%	80.5% / 2.0%
OMA	VOC (F-RCNN)	79.3%	44.9% / 94.6%	78.1% / 17.4%
OMA	COCO (YOLO)	52.8%	0.4% / 94.6%	52.0% / 1.5%
OMA	COCO (F-RCNN)	49.7%	6.3% / 91.9%	47.0% / 16.3%
ODA	VOC (YOLO)	72.0%	71.6% / 96.5%	76.7% / 20.8%
ODA	VOC (F-RCNN)	77.6%	76.4% / 69.3%	75.4% / 18.9%
ODA	COCO (YOLO)	54.0%	52.4% / 99.9%	54.1% / 25.4%
ODA	COCO (F-RCNN)	50.9%	50.3% / 81.7%	51.4% / 28.0%
OGA	VOC (YOLO)	80.4%	78.0% / 65.1%	82.2% / 0.0%
OGA	VOC (F-RCNN)	83.2%	81.2% / 98.4%	80.9% / 0.0%
OGA	COCO (YOLO)	53.0%	52.8% / 99.8%	54.4% / 0.0%
OGA	COCO (F-RCNN)	48.9%	49.1% / 95.4%	49.5% / 0.0%

Poké Ball	OMA	77.3%	95.5%	19.8%
Poké Ball	ODA	75.3%	98.5%	35.1%
Poké Ball	OGA	79.8%	96.8%	15.4%
Solid White	OMA	75.5%	82.0%	52.9%
Solid White	ODA	71.8%	71.1%	44.0%
Solid White	OGA	80.2%	73.7%	37.0%

ODPure

Most Popular