Photo Response Non-Uniformity Based AI-Generated Image Detection

The hyper-realism of images generated by diffusion models has become increasingly impressive, making it difficult for humans to distinguish between AI-generated and real images. Most existing detectors rely on extracting artifacts inherent in the process of creating fake images, often ignoring the unique fingerprints left behind when real images are captured. It has been observed that real images exhibit a distinctive pattern noise known as Photo Response Non-Uniformity (PRNU). Based on this consideration, in this study, an Al-generated image detection method based on PRNU is proposed. Specifically, by combining PRNU features with original image features, we effectively capture the intrinsic noise patterns in real images and then train a classification network. Experimental results on 11 generative models demonstrate the practical effectiveness of this method in detecting AI-generated images.