Breaking Semantic Artifacts for Generalized AI-generated Image Detection

This repository contains code for Breaking Semantic Artifacts for Generalized AI-generated Image Detection (https://nips.cc/virtual/2024/poster/95403, NeurIPS 2024) by Chende Zheng, Chenhao Lin∗, Zhengyu Zhao, Hang Wang, Xu Guo, Shuai Liu∗, Chao Shen.

Environment Setup

Our codebase requires the following Python:

• Python >= 3.8.17
• PyTorch >= 1.13.1

You can set up the environment by following these steps:

1. Clone this repository

   git clone https://github.com/Zig-HS/FakeImageDetection.git
   cd FakeImageDetection
   

2. Install the necessary libraries

   pip install -r requirements.txt
   

Data Prepare

• You can prepare your data through the following directory structure:

  datasets
  └── train					
        ├── gan
        │── dm   	
        │     .
        │     .
  └── val
        ├── gan
        │── dm   	
        │     .
        │     .
  └── test
        ├── progan
        │── cyclegan   	
        │     .
        │     .
  

• Each directory (e.g., gan, progan) will contain real/fake images under 0_real and 1_fake folders respectively.

Training

• After prepareing the data, you can train the model with the following command:

  python train.py -j options/default.json
  

Evaluation

• You can evaluate the model on all the dataset included in ./datasets/test or other folders by running:

  python eval.py -j default.json -r datasets/test
  

• The results will be stored in results/<model_name>.txt, representing the Average Prevision and AUROC scores for each of the test domains.