Detection Enhancement of AI-Synthesized Images by Fusing High-Level GAN-Based Features and Low-Level Texture Features

As the quality of fake image generation continues to improve, concerns have already begun to arise regarding the impact of the spread of such falsities. Although existing detection methods have performed well in detecting generated images, their generalization capability in real-world scenarios is poor. This paper introduces existing detection methods and their limitations, finding that simply using one type of feature may no longer be effective in detecting current AI synthetic images, and thus proposes the fusion of multiple features to enhance the detection capability of the detector. Specifically, we extract the high-frequency regions and gradient images of the images as low-level and high-level features for detection, respectively, and then fuse the low-level and high-level features through a fusion module based on the multi-head attention mechanism to improve the accuracy and generalization capability of detection. Our method has been extensively tested on datasets generated by 10 different models, and the experimental results demonstrate that our approach exceeds the performance of current leading detectors in identifying AI-generated synthetic images, whether they originate from known or unknown sources.