Search papers, labs, and topics across Lattice.
This paper investigates the limitations of AI-generated images in replicating the color statistics of natural images, revealing that the LPIPS loss function is less sensitive to chrominance. By introducing six hand-crafted color transformations and a learned task-optimized color transform, the authors develop a robust method for distinguishing synthetic images from real ones, achieving an impressive average generalization accuracy of 93.27%. The findings highlight the potential of color statistics as a reliable forensic tool in the ongoing battle against AI-generated image deception.
Color statistics expose a critical vulnerability in AI-generated images, enabling a 93.27% accuracy in distinguishing them from real ones.
The evolution and dissemination of AI-synthesized images is occurring at an unprecedented rate. Image generators are making rapid progress in their goal of perfectly imitating natural images, which also challenges image forensics. In this work, we exploit an underexplored cue in current generative models, namely their weakness to imitate color statistics of natural images. We first show that the LPIPS loss used for training image generators is less sensitive to chrominance than to luminance, which may lead to statistical discrepancies in the colors of synthetic images. Building on this observation, we then introduce six hand-crafted color transformations and a method to learn a task-optimized color transform to statistically expose generated images. These transformations can be used in various ways. First, we define color-sensitive features at pixel-level or patch-level. A simple, interpretable classifier achieves with these features an average generalization accuracy of 93.27% and strong robustness against six types of post-processing. Second, we demonstrate that the transformations exhibit characteristic visual noise patterns in natural and synthetic image areas, which enables an intuitive visual image evaluation. Third, we demonstrate that the transforms can enhance color patterns in generated images for improved multiclass attribution.
