베이어 패턴 아티팩트 검출을 통한 AI 생성 이미지의 진위 판별

초록

The rapid advancement of artificial intelligence(AI) based image generation technologies has reached a point where AI-generated images are virtually indistinguishable from real photographs. This progression has significantly undermined the reliability of images with uncertain origins as credible evidence for authenticity verification, leading to pressing societal issues such as the spread of misinformation and the proliferation of fake news. In response to these challenges, this thesis introduces a novel analytical method grounded in the physical characteristics of camera-captured images, Re-interpolation techniques to detect subtle differences between authentic and AI-generated images. The proposed method involves a comprehensive analysis of pixel patterns inherent in camera-captured images, followed by the application of Re-interpolation techniques to effectively distinguish between the two image types. Experimental results reveal statistically significant differences between camera-captured and AI-generated images, thereby affirming the accuracy of the proposed method in determining image authenticity. Furthermore, the method developed in this study is anticipated to markedly enhance the reliability of digital images and contribute to the prevention of misinformation and the dissemination of fake news. The outcomes of this research suggest that the proposed method holds substantial practical potential for the verification of digital content authenticity. Additionally, future research will focus on further validation and refinement of this method to extend its applicability to a broader spectrum of image generation technologies.

키워드