MaskSim: Detection of synthetic images by masked spectrum similarity analysis

Synthetic image generation methods have recently revolutionized the way in which visual content is created. This opens up creative opportunities but also presents challenges in preventing misinformation and crime. However, these methods leave traces in the Fourier spectrum that are invisible to humans, but can be detected by specialized tools. This paper describes a semi-white-box method for detecting synthetic images by revealing anomalous patterns in the spectral domain. Specifically, we train a mask to enhance the most discriminative frequencies and simultaneously train a reference pattern that resembles the patterns produced by a given generative method. The proposed method produces explainable results with state-of the-art performances and highlights cues that can be used as forensic evidence. Code is available at https://github.com/li-yanhao/masksim.