DADA

Targeted False Positive Synthesis via Detector-guided Adversarial Diffusion Attacker for Robust Polyp Detection

Quan Zhou^{1, *}, Gan Luo^{1, *}, Qiang Hu^{2, †}, Qingyong Zhang¹, Jinhua Zhang³, Yinjiao Tian³, Qiang Li², Zhiwei Wang^{2, †}

¹ WUT, ² WNLO, HUST, ³ Changzhou United lmaging Healthcare Surgical Technology Co., Ltd.
(*: equal contribution, †: corresponding author)

Overview

1. Inference Flow

In our work, we adopt a inpainting strategy and the specific inference flow is: Firstly you should have a real endoscopy image; Secondly you need to specify a inpainting region (a bounding box region); and Finally, our method can generate a high-value negative sample in the region you have formulated in this real image. To make it clear, we've made a GIF, which is shown as follows:

2. Introduction

To tackle the challenge of polyp detectors easy to produce false positive prediction, this work firstly (to the best of our knowledge) to apply adversarial diffusion framwork to generate high-value negative samples (e.g., polyp-like interferences, light spots, circular lumens and and other background objects). The core module is Detector-guided Adversarial Diffusion Attacker (DADA), which perturbs the denoising process to generate challenging false positives capable of confusing a well-trained polyp detector, thereby guiding the synthesis of high-value negative samples. The negative samples synthesized by our method can improve the performance of the detector more than other data data synthesis methods.

Experimental Results

1. Comparison with State-of-the-arts

- Quantitative Comparisons

- Qualitative Comparisons

Quick Start

- Preliminaries

• Python 3.8+
• PyTorch 2.2+
• TorchVision corresponding to the PyTorch version
• NVIDIA GPU + CUDA

1. Install dependencies.

cd DADA
pip install -r requirements.txt

2. Data synthesis

We provide well-trained BG-De weights based on the public Kvasir dataset, which can be downloaded from this link, . Please place the weights in the BG-De_model folder.

Additionally, we also offer YOLOv5l weights trained on the public Kvasir dataset, which can be downloaded from this link. Please place the weights in the Detection_model folder.

  DADA
  ├── BG-De_model
  ├── Detection_model

After completing the above steps, you can generate negative samples by simply running Main.py:

python Main.py

Training

1. Prepare the datasets for BG-De.

Please place the dataset you want to train in the path ./datasets and ensure that the size of the images and masks is 256. The path structure should be as follows:

  DADA
  ├── datasets
  │   ├── images
  │   ├── masks

2.Train your own BG-De.

Please set the "state" parameter of modelConfig in Main.py to "train", and set parameters such as batch_size according to actual conditions.

python Main.py

Place the weights in the BG-De_model folder.

- Testing

Before running the tests, ensure an object detection model is ready. For this project, we utilize the YOLOv5l architecture. Place the weights in the Detection_model folder and set the "state" parameter of modelConfig in Main.py to "eval".

python Main.py

Acknowledgments

We acknowledge the outstanding open-source contributions from DDPM, YOLOv5, and DETR.

Citation

If you find our paper and code useful in your research, please consider giving us a star ⭐ and citing DADA by the following BibTeX entry.

@article{zhou2025targeted,
  title={Targeted False Positive Synthesis via Detector-guided Adversarial Diffusion Attacker for Robust Polyp Detection},
  author={Zhou, Quan and Luo, Gan and Hu, Qiang and Zhang, Qingyong and Zhang, Jinhua and Tian, Yinjiao and Li, Qiang and Wang, Zhiwei},
  journal={arXiv preprint arXiv:2506.18134},
  year={2025}
}