跳转至

文章背景与核心概要

随着生成式人工智能的爆发式增长，业界迫切需要强大的合成图像检测方法。当前的研究主要集中在“跨生成器”的泛化能力上，而本文指出，“跨域（cross-domain）”泛化能力同样至关重要。为了解决这一问题，作者推出了 Fusion`Detect 这一创新框架，它利用了来自两个冻结基础模型（CLIP 和 Dinov2）的内聚特征空间。通过结合这些互补模型，FusionDetect 对内容和生成器设计都实现了极佳的适应性。

此外，该研究还引入了 OmniGen Benchmark（全能生成基准），这是一个包含 12 个先进生成器的综合数据集，用于在现实且多样化的条件下评估检测器的性能。该论文在准确率和精确度上均超越了现有模型，展现出对抗常见图像扰动的卓越鲁棒性。

文章背景与核心概要

The proliferation of generative AI has necessitated robust methods for synthetic image detection. While current research primarily focuses on "cross-generator" generalization, this paper argues that "cross-domain" generalization is equally critical. To address this, the authors introduce FusionDetect, a novel framework that leverages a cohesive feature space derived from two frozen foundation models—CLIP and Dinov2. By combining these complementary models, FusionDetect achieves superior adaptability to both content and generator design. The study also introduces the OmniGen Benchmark, a comprehensive dataset featuring 12 state-of-the-art generators, to evaluate detector performance under realistic, diverse conditions.

Redefining Generalization in Visual Domains: ATwo-Axis Framework for Fake Image Detection with FusionDetect

Redefining Generalization in Visual Domains: A Two-Axis Framework for Fake Image Detection with FusionDetect

Authors: Amirtaha Amanzadi, Zahra Dehghanian, Hamid Beigy, Hamid R. Rabiee
arXiv: 2510.05740 [cs.CV]
Submitted: 7 Oct 2025 (v1), 14 Aug 2026 (v2)

Authors: Amirtaha Amanzadi, Zahra Dehghanian, Hamid Beigy, Hamid R. Rabiee
arXiv: 2510.05740 [cs.CV]
Submitted: 7 Oct 2025 (v1), 14 Aug 2026 (v2)

Summary

Summary

生成式 AI 的泛滥使得构建可靠的合成图像检测方法变得必不可少。尽管目前的研究主要集中在“跨生成器”（cross-generator）的泛化能力上，但本文认为“跨域”（cross-domain）泛化同样至关重要。为了应对这一挑战，作者引入了 FusionDetect，这是一个新颖的框架，它利用了来自两个冻结基础模型（CLIP 和 Dinov2）的内聚特征空间。通过结合这些互补的模型，FusionDetect 对内容和生成器设计均实现了卓越的适应性。该研究还推出了 OmniGen Benchmark，这是一个包含 12 个最先进生成器的综合数据集，用于评估检测器在真实且多样化条件下的性能。

The proliferation of generative AI has necessitated robust methods for synthetic image detection. While current research primarily focuses on "cross-generator" generalization, this paper argues that "cross-domain" generalization is equally critical. To address this, the authors introduce FusionDetect, a novel framework that leverages a cohesive feature space derived from two frozen foundation models—CLIP and Dinov2. By combining these complementary models, FusionDetect achieves superior adaptability to both content and generator design. The study also introduces the OmniGen Benchmark, a comprehensive dataset featuring 12 state-of-the-art generators, to evaluate detector performance under realistic, diverse conditions.

Key Contributions

Key Contributions

  • 双轴泛化框架： 作者通过同时解决跨生成器和跨域泛化问题，重新定义了虚假图像检测的问题。
  • FusionDetect 方法： 一种全新的检测架构，它融合了 CLIP 和 Dinov2 的特征，构建出一个强大且统一的特征空间。
  • OmniGen 基准： 一个整合了 12 个最先进生成模型的新评估标准，用于在真实世界场景中测试检测器的性能。
  • 最先进的性能表现：
    • 在现有基准上，准确率提升了 3.87%，精确度提升了 6.13%。
    • 在全新的 OmniGen 基准上，准确率提升了 4.48%。
    • 展现出对常见图像扰动的杰出鲁棒性。
  • Two-Axis Generalization Framework: The authors redefine the problem of fake image detection by addressing both cross-generator and cross-domain generalization.
  • FusionDetect Method: A new detection architecture that fuses features from CLIP and Dinov2 to create a robust, unified feature space.
  • OmniGen Benchmark: A new evaluation standard incorporating 12 state-of-the-art generative models to test detector performance in real-world scenarios.
  • State-of-the-Art Performance:
    • Outperforms existing models by 3.87% in accuracy and 6.13% in precision on established benchmarks.
    • Achieves a 4.48% increase in accuracy on the new OmniGen benchmark.
    • Demonstrates exceptional robustness against common image perturbations.

Subjects

Subjects

  • 主分类： 计算机视觉与模式识别 (cs.CV)
  • 次分类： 人工智能 (cs.AI)
  • Primary: Computer Vision and Pattern Recognition (cs.CV)
  • Secondary: Artificial Intelligence (cs.AI)