MSAFNet: multi-scale self-adaptive feature fusion network for AI-generated image detection

Abstract With the development of advanced generative technologies such as generative adversarial networks and diffusion models, AI-generated images have become extremely realistic. As a result, there is an urgent need for efficient detection tools to combat misinformation and protect societal trust and personal privacy. Current detection methods perform well when identifying content from known generative models, but their performance drops when facing unknown or emerging technologies. Moreover, most detection methods focus on detecting local artifacts, overlooking the extraction and effective fusion of global relationships and multi-scale features, which makes it difficult to capture the complex patterns in AI-generated images. To address these challenges, this paper proposes a multi-scale self-adaptive feature fusion network (MSAFNet). Specifically, the proposed MSAFNet utilizes a global noise feature extraction module to capture global noise patterns in images, compensating for the shortcomings of existing methods that focus solely on local noise features. Meanwhile, it adopts a texture feature extraction module to capture subtle texture anomalies and a color feature extraction module to enrich color feature information, enhancing the expression capability of color features and improving the recognition of complex features in AI-generated images. Lastly, we introduce a self-adaptive feature fusion module to capture complementary information and the importance of multi-scale features, ensuring efficient fusion of feature information across different dimensions. Experimental results on the large-scale DFFD and CIFAKE datasets demonstrate that the proposed MSAFNet achieves higher classification accuracy compared to other AI-generated image detection methods.