This project addresses the challenge of reliably detecting AI-generated images, particularly those exhibiting subtle artifacts or designed to evade current detection methods. It targets researchers and practitioners in computer vision and AI ethics, providing a more rigorous benchmark and a novel detection model to assess the true state of AI-generated image detection. The benefit lies in a more accurate understanding of detection capabilities and a tool to combat sophisticated AI image generation.

How It Works

The project introduces the Chameleon dataset, comprising AI-generated images crafted to be genuinely challenging for human perception and existing detection algorithms. It then performs a "sanity check" by evaluating nine off-the-shelf detectors on this dataset, revealing significant failure rates where models misclassify AI-generated images as real. To improve detection, the AIDE (AI-generated Image Detector with Hybrid Features) model is proposed, which leverages multiple experts to simultaneously extract diverse visual artifacts and noise patterns, aiming for enhanced generalization and robustness.

Quick Start & Requirements

• Installation: Clone the repository, set up a conda environment with PyTorch 2.0.1, torchvision 0.15.2, and torchaudio 2.0.2, then install remaining packages via pip install -r requirements.txt.
• Prerequisites: CUDA 11.8, Python 3.10, PyTorch 2.0.1. Pre-trained ResNet and ConvNeXt models are required for training.
• Links: GitHub repository: https://github.com/shilinyan99/AIDE
• Dataset Access: The Chameleon dataset requires emailing tattoo.ysl@gmail.com.

Highlighted Details

• Presents the Chameleon dataset, designed to simulate challenging real-world scenarios for AI-generated image detection.
• Demonstrates that most existing AI-generated image detectors fail to generalize to this challenging dataset, classifying AI images as real.
• Proposes AIDE, a hybrid feature detector utilizing multiple experts for artifact and noise pattern extraction.

Maintenance & Community

Limited community interaction details are provided. For inquiries regarding the project or the Chameleon dataset, contact tattoo.ysl@gmail.com.

Licensing & Compatibility

The Chameleon dataset is strictly for academic research use and prohibits commercial use. The license for the AIDE code itself is not explicitly stated in the provided text.

Limitations & Caveats

The primary caveat highlighted is the significant underperformance of current state-of-the-art AI-generated image detection models on challenging, realistic datasets, suggesting the task is far from being definitively "solved." Access to the Chameleon dataset requires direct email contact.