Detecting AI Generated Images Through Texture and Frequency Analysis of Patches

The significant improvement in AI image generation in recent years poses serious threats to social security, as AI generated misinformation may infringe upon political stability, personal privacy, and digital copy rights of artists. Building an AI generated image detector that accurately identifies generated image is crucial to maintain the social security and property rights of artists. This paper introduces preprocessing pipeline that uses positional encoded azimuthal integrals for image patches to create fingerprints that encapsulate distinguishing features. We then trained a multi-head attention model with 97.5% accuracy on classification of the fingerprints. The model also achieved 80% accuracy on images generated by AI models not presented in the training dataset, demonstrating the robustness of our pipeline and the potential of broader application of our model.