Where’s the liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content

This repository contains the source code for Where’s the liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content.

Overview

The recent proliferation of photorealistic images created by generative models has sparked both excitement and concern, as these images are increasingly indistinguishable from real ones to the human eye. While offering new creative and commercial possibilities, the potential for misuse, such as in misinformation and fraud, highlights the need for effective detection methods. Current detection approaches often rely on access to model weights or require extensive collections of real image datasets, limiting their scalability and practical application in real-world scenarios. In this work, we introduce a novel black-box detection framework that requires only API access, sidestepping the need for model weights or large auxiliary datasets.

Setups

1. Clone this repository

git clone https://github.com/HaoyueBaiZJU/genai-detect
cd genai-detect

2. Install the necessary libraries

To set up the required environment, install dependencies from requirements.txt:

pip install -r requirements.txt
pip install git+https://github.com/cloneofsimo/lora.git

Quick Start

Data Preparation:

• Dataset for the diffusion models (e.g., LDM/Glide) can be found here.

• Download and unzip the file into the following directory.

./real_folder

./fake_folder

./masks_folder

The different types of masks data can be downloaded vis the link:

https://drive.google.com/file/d/1xlnTxxO0CC1JD--MgLplaIYVOF_bQ_Am

The dalle3-pop and corresponding genhalf masks example can be downloaded via the link:

https://drive.google.com/file/d/1l3vMHdPgeKXK-MlB8e3Sd1MTtKaTt5GM

Running the Evaluation

To evaluate a model on the dataset, run the following command:

python eval_real_vs_fake.py \
    --real_folder /path/to/real_images \
    --fake_folder /path/to/fake_images \
    --seeds 0 \
    --save_location ./ckpt \
    --checkpoint_folder ./my_sd_inpainting_ckpt \
    --num_inference_steps 50 \
    --guidance_scale 7

You can test the benchmark results using the script:

python evaluate_scores.py

To conduct fine-tuning via lora, run the following command:

bash inpainting_lora.sh

Citation

If you use our codebase, please cite our work:

@inproceedings{bai2025s,
  title={Where's the Liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content},
  author={Bai, Haoyue and Sun, Yiyou and Cheng, Wei and Chen, Haifeng},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={28821--28830},
  year={2025}
}