NSFF: Noise and semantic features fusion for AI-generated image detection

Advanced generative models are capable of producing photorealistic images with strong semantic consistency and a high degree of controllability. Such images are highly convincing in real-world scenarios and may facilitate the malicious spread and misuse of misinformation. Therefore, it is urgently needed to develop reliable detectors to identify AI-generated images. However, previous methods have generally relied on large-scale training sets composed of real and generated images, which not only lead to lengthy training processes but also often suffer from significant performance degradation when detecting images produced by unknown generative models. To address this issue, we propose a general detection framework for AI-generated images, termed NSFF. Unlike previous methods that primarily combined texture features and semantic information, NSFF explicitly utilizes noise artifacts in poorly textured regions and combines global and local semantic information to obtain more complementary and robust representations. We extracted an equal number of real and AI-generated images from four benchmark datasets to construct the testing set. The experiments show that NSFF can effectively generalize to images generated by previously unseen models with only 2000 images for small-scale supervised training. In the test set, our method achieved an average accuracy of 80.82%. Our code can be found at https://github.com/IMUT-YangHr/NSFF-main.