iLearn-Lab/TDSC26-ERASE

🕵️‍♂️ ERASE: Bypassing Collaborative Detection of AI Counterfeit (Model Weights)

Qianyun Yang¹  Peizhuo Lv²  Yingjiu Li³  Shengzhi Zhang⁴  Yuxuan Chen¹  Zixu Li¹  Yupeng Hu¹

¹Shandong University  ²Nanyang Technological University  ³University of Oregon   ⁴Boston University

These are the official pre-trained model weights for **ERASE**, an optimization framework designed to bypass single and collaborative detection of AI-Generated Images (AIGI) by comprehensively eliminating multi-dimensional generative artifacts.

🔗 Paper: [Accepted by IEEE TDSC 2026] (Coming Soon)
🔗 GitHub Repository: iLearn-Lab/TDSC26-ERASE

---

📌 Model Information

1. Model Name

ERASE (comprehensive counterfeit ArtifactS Elimination) Checkpoints.

2. Task Type & Applicable Tasks

• Task Type: Adversarial Attack / AI-Generated Image Stealth (AIGI-S) / Image-to-Image
• Applicable Tasks: Bypassing AI-generated image detectors (both single detectors and collaborative multi-detector environments) while maintaining exceptionally high visual fidelity.

3. Project Introduction

With the rapid development of generative AI, the issue of deepfakes has become increasingly severe. Existing AI-Generated Image Stealth (AIGI-S) methods typically optimize against a single detector and often fail when facing real-world "Collaborative Detection". Moreover, they often introduce obvious artifacts visible to human observers.

ERASE is a stealth optimization framework that innovatively combines:

• 🎯 Sensitive Feature Attack
• ⛓️ Diffusion Chain Attack (Optimization-free)
• 📻 Decoupled Frequency Domain Processing

This Hugging Face repository hosts the pre-trained weights required to run the Decoupled Frequency Domain Processing and the Surrogate Classifiers, specifically noise_prototype_VAE.pt, dncnn_color_blind.pth, and the ckpt_ori surrogate weights.

4. Training Data Source

The surrogate classifiers and related components were primarily trained and evaluated on the GenImage dataset, following the standard task settings of AIGI-S evaluation.

---

🚀 Usage & Basic Inference

These weights are designed to be used seamlessly out-of-the-box with the official ERASE GitHub repository.

Step 1: Prepare the Environment

Clone the GitHub repository and install dependencies:

git clone https://github.com/iLearn-Lab/TDSC26-ERASE
cd ERASE
conda create -n erase python=3.9 -y
conda activate erase
pip install -r requirements.txt

Step 2: Download Model Weights

Download the files from this Hugging Face repository (ckpt_ori folder, noise_prototype_VAE.pt, dncnn_color_blind.pth) and place them in the checkpoints/ directory of your cloned GitHub repo. Your structure should look like this:

ERASE/
└── checkpoints/
    ├── ckpt_ori/                 # Surrogate model weights (E/R/D/S)
    ├── noise_prototype_VAE.pt    # Frequency VAE weights
    └── dncnn_color_blind.pth     # Denoising/Frequency weights

Step 3: Run the Attack

Use main.py from the code repository to perform basic inference and generate adversarial images:

python main.py \
    --images_root ./input_images \
    --save_dir ./output \
    --model_name E,R,D,S \
    --diffusion_steps 20 \
    --start_step 18 \
    --iterations 10 \
    --is_encoder 1 \
    --encoder_weights ./checkpoints/noise_prototype_VAE.pt \
    --eps 4 \
    --batch_size 4 \
    --device cuda:0

---

⚠️ Limitations & Notes

Disclaimer: This tool and its associated model weights are strictly intended for academic research, AI security evaluation, and robustness testing.

• It is strictly prohibited to use this repository for any malicious forgery, fraud, or other illegal/unethical purposes.
• Users bear full legal responsibility for any consequences arising from improper use.

---

📝⭐️ Citation

If you find our weights or code useful for your research, please consider leaving a Star ⭐️ on our GitHub repo and citing our paper:

@article{yang2026erase,
  title={ERASE: Bypassing Collaborative Detection of AI Counterfeit via Comprehensive Artifacts Elimination},
  author={Yang, Qianyun and Lv, Peizhuo and Li, Yingjiu and Zhang, Shengzhi and Chen, Yuxuan and Chen, Zhiwei and Li, Zixu and Hu, Yupeng},
  journal={IEEE Transactions on Dependable and Secure Computing},
  year={2026},
  publisher={IEEE}
}