AI-Generated Image Detection Using Semantic Feature
Hau-Ching Chen, Yu-Pin Shen, Yi-Hung Yeh, Jiann-Liang Chen · 2025
In recent years, the rapid advancement of deep learning has brought generative models to the forefront of technological innovation, enabling the creation of highly realistic images from simple text inputs. However, this progress has also introduced new challenges in distinguishing real from synthetic content, a concern particularly evident in the misuse of such technologies for malicious purposes. This paper investigates the generalization capabilities of the CLIP Vision Transformer (ViT) model for fake image detection, leveraging its semantic understanding of images. We employ a cross-validation approach using the GenImage dataset to evaluate the performance of CLIP ViT against traditional models like ResNet-50 and Swin-T. Results demonstrate that CLIP ViT exhibits superior generalization ability on unseen datasets, outperforming other models by focusing on high-level semantic features rather than dataset-specific patterns. This study highlights the potential of CLIP ViT in advancing the detection of generative model outputs by capturing underlying semantic relationships, suggesting a paradigm shift towards more robust and generalized detection techniques.