Distinguishing AI-Generated and Real Images Using Swin-Transformer

The recent advancements in generative AI technology have enabled the effortless creation of realistic synthetic images and videos, exacerbating issues related to privacy breaches and the generation of inappropriate content. Such synthetic content can lead to various societal challenges, including violations of privacy, defamation, and the dissemination of false information, necessitating urgent development of technical solutions for detection and prevention. This study proposes a novel detection model leveraging the shifted window mechanism and hierarchical structure of the Swin-Transformer to significantly enhance the capability to distinguish between AI-generated and real images. The proposed model was trained and validated using the RVF10k and Real and Fake Face Detection datasets to evaluate its generalization performance across diverse types of synthetic images. The model's detection capabilities were further strengthened by employing preprocessing and augmentation techniques such as Gaussian Blurring, Canny Edge Detection, Sobel Filter, and Laplacian Filter. Experimental results demonstrate that the proposed model outperforms existing CNN and Transformer-based models, reliably detecting even expertly manipulated synthetic images. This study establishes a robust technical foundation to address the misuse of generative AI, aiming to further enhance the model's generalization capabilities through future expansions to various synthetic media types and the development of lightweight models capable of real-time detection. By doing so, it seeks to provide effective countermeasures against privacy breaches and the creation of inappropriate content caused by generative AI, promoting its safe utilization and contributing to the restoration of societal trust.