Scholarlink

<p>The exponential advancement in generative artificial intelligence (AI) and physically based rendering has severely blurred the perceptual boundary between authentic photographs and computer-generated imagery (CGI). The primary purpose of this comprehensive survey is to critically evaluate deep learning-based methodologies dedicated to CGI detection and visual forgery analysis. As our core methodology, we propose a systematic taxonomy that categorizes the existing literature into three distinct technical domains: architectural paradigm shifts (from convolutional neural networks—CNNs—to vision transformers), multi-domain feature representations (spatial, spectral, and color spaces), and multimodal fusion strategies. A rigorous quantitative synthesis of reported outcomes exposes a critical generalization bottleneck: While state-of-the-art models routinely achieve &gt; 95% detection accuracy in isolated, intra-dataset evaluations (e.g., DSTok, FaceForensics++), they suffer severe accuracy degradation under cross-generator testing and real-world compressions. To overcome these critical limitations, this study identifies existing methodological gaps and outlines high-potential future research trajectories—specifically emphasizing hybrid multi-color space architectures, multitask learning, and the integration of explainable AI (XAI). Ultimately, this review establishes a foundational framework to guide the development of robust, highly generalizable digital forensic systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning for CGI and visual forgery detection: a comprehensive survey

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Ahmet Korkmaz,
  • Mustafa Özden

摘要

The exponential advancement in generative artificial intelligence (AI) and physically based rendering has severely blurred the perceptual boundary between authentic photographs and computer-generated imagery (CGI). The primary purpose of this comprehensive survey is to critically evaluate deep learning-based methodologies dedicated to CGI detection and visual forgery analysis. As our core methodology, we propose a systematic taxonomy that categorizes the existing literature into three distinct technical domains: architectural paradigm shifts (from convolutional neural networks—CNNs—to vision transformers), multi-domain feature representations (spatial, spectral, and color spaces), and multimodal fusion strategies. A rigorous quantitative synthesis of reported outcomes exposes a critical generalization bottleneck: While state-of-the-art models routinely achieve > 95% detection accuracy in isolated, intra-dataset evaluations (e.g., DSTok, FaceForensics++), they suffer severe accuracy degradation under cross-generator testing and real-world compressions. To overcome these critical limitations, this study identifies existing methodological gaps and outlines high-potential future research trajectories—specifically emphasizing hybrid multi-color space architectures, multitask learning, and the integration of explainable AI (XAI). Ultimately, this review establishes a foundational framework to guide the development of robust, highly generalizable digital forensic systems.