RealStats：一种用于假图像检测的严格实数统计框架

As generative models continue to evolve, detecting AI-generated images remains a critical challenge. While effective detection methods exist, they often lack formal interpretability and may rely on implicit assumptions about fake content, potentially limiting robustness to distributional shifts. This work introduces a rigorous, statistically grounded framework for fake image detection that focuses on producing an interpretable probability score. The RealStats methodology aims to enhance detection transparency and reliability by exclusively utilizing real-valued statistical data, thereby avoiding prior assumptions about fabricated content. This framework distinguishes between real and synthetic images by analyzing the intrinsic statistical properties of image data, rather than relying on generator-specific artifacts or traces. The core idea is to construct a series of statistical tests capable of quantifying an image's 'realness' or 'non-realness,' aggregating these into a unified probabilistic score. The advantage of this methodology lies in its strong theoretical foundation and its ability to provide a quantifiable, easily understandable confidence level indicating whether an image is authentic or AI-generated. The framework's design considers robustness against various generative models and potential attacks, aiming to address future advancements in generation technology. By avoiding specific assumptions about the generation process, RealStats strives to offer a more universal and enduring solution to the evolving needs of AI-generated content detection. Experimental evaluations of this framework will demonstrate its effectiveness and generalization capabilities across different datasets and generative models.