Scholarlink

<p>The rapid proliferation of advanced image generation models, particularly diffusion models, poses significant threats to digital security. Existing single-domain detection methods often struggle to discern the subtle artifacts introduced by these emerging AI generations. In this study, we propose a specialized dual-stream framework that strategically integrates features from both spatial and frequency domains to disentangle the fingerprints of heterogeneous generators. We address two critical tasks: binary real/fake detection and closed-set source attribution across 10 distinct generative architectures. By conducting extensive experiments with multiple backbones (ResNet-50, EfficientNet-B3, and ViT) on a diverse dataset of 15 classes, we demonstrate that our fusion strategy significantly enhances discriminative power. Empirical results identify EfficientNet-B3 as the optimal backbone, achieving 96.04% accuracy for binary classification and 95.92% for attribution, surpassing single-domain baselines. Finally, we employ XGrad-CAM to demonstrate that the model correctly focuses on technical fingerprints and spectral anomalies rather than semantic content, thereby reinforcing the system’s reliability for practical applications.</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

Dual-stream framework for real/fake classification and source attribution of AI-generated images using spatial and frequency features

  • 期刊论文

  • 发表日期：

    2026年7月15日

查看全文
  • Linh Thuy Thi Pham,
  • Cu Vinh Loc,
  • Truong Nhat Tran,
  • Hai Thanh Nguyen

摘要

The rapid proliferation of advanced image generation models, particularly diffusion models, poses significant threats to digital security. Existing single-domain detection methods often struggle to discern the subtle artifacts introduced by these emerging AI generations. In this study, we propose a specialized dual-stream framework that strategically integrates features from both spatial and frequency domains to disentangle the fingerprints of heterogeneous generators. We address two critical tasks: binary real/fake detection and closed-set source attribution across 10 distinct generative architectures. By conducting extensive experiments with multiple backbones (ResNet-50, EfficientNet-B3, and ViT) on a diverse dataset of 15 classes, we demonstrate that our fusion strategy significantly enhances discriminative power. Empirical results identify EfficientNet-B3 as the optimal backbone, achieving 96.04% accuracy for binary classification and 95.92% for attribution, surpassing single-domain baselines. Finally, we employ XGrad-CAM to demonstrate that the model correctly focuses on technical fingerprints and spectral anomalies rather than semantic content, thereby reinforcing the system’s reliability for practical applications.