文献类型：期刊文献

英文题名：Improved Model in generated Image Detection

作者：Chen, Zihan[1]; Yin, Liangtao[1]; Du, Yujie[1]; Wu, Shengxi[1]

机构：[1] School of Information Science and Technology, East China University of Science & Technology, Shanghai, China

年份：2025

卷号：13815

外文期刊名：Proceedings of SPIE - The International Society for Optical Engineering

收录：EI(收录号：20255219764421)

语种：英文

外文关键词：Dimensionality reduction - Error detection - Face recognition - Feature extraction - Image recognition - Image texture - Modal analysis - Semantics - Textures - Visualization

摘要：The rapid advancement of generative AI technologies has enabled the creation of hyper-realistic facial images that compromise conventional detection paradigms, posing critical risks to privacy preservation and authentication systems. To address this challenge, an improved detection model architecture integrating PatchCraft and SigLip is proposed, overcoming the limitations of single-modality approaches by achieving comprehensive recognition of both texture and semantic features through synergistic combination of texture-based artifact detection with cross-modal semantic analysis. This approach is innovatively applied to the identification of AI-generated facial images. Through UMAP-based dimensionality reduction visualization and Grad-CAM heatmap analysis, we systematically validate the model's learned feature representations. Experimental results demonstrate that the proposed architecture achieves superior accuracy and robustness, significantly outperforming conventional models in AI-generated facial image recognition tasks. ? 2025 SPIE.