数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (1): 68-85.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.01.005

doi: 10.11871/jfdc.issn.2096-742X.2025.01.005

• 专刊：生成式人工智能 • 上一篇    下一篇

基于集成学习的AI生成图像对抗检测框架

金维正(),唐云祁*()   

  1. 中国人民公安大学侦查学院，北京 100038
  • 收稿日期:2024-10-16 出版日期:2025-02-20 发布日期:2025-02-21
  • 通讯作者: *唐云祁（E-mail: tangyunqi@ppsuc.edu.cn）
  • 作者简介:金维正，中国人民公安大学刑事科学技术专业硕士研究生，主要研究方向为图像检测。
    负责实验设计和实现以及论文初稿撰写。
    JIN Weizheng is currently a master’s student in the Forensic Science, at the People’s Public Security University of China. His research interests include image detection
    In this paper, he is responsible for experimental design and implementation, as well as writing the initial draft of the paper.
    E-mail: 2022211383@stu.ppsuc.edu.cn|唐云祁，教授，博士生导师，主要研究方向为人工智能、刑事科学技术等。
    负责论文框架设计，论文修改与审定。
    TANG Yunqi, is a professor and doctoral supervisor at the People’s Public Security University of China. His research interests include artificial intelligence, forensic science, etc.
    In this paper, he is responsible for designing the paper framework, revising and approving the paper.
    E-mail: tangyunqi@ppsuc.edu.cn
  • 基金资助:
    中国人民公安大学刑事科学技术双一流创新研究专项(2023SYL06)

Adversarial Detection Framework for AI-Generated Images Based on Ensemble Learning

JIN Weizheng(),TANG Yunqi*()   

  1. School of Investigation, People’s Public Security University of China, Beijing 100038, China
  • Received:2024-10-16 Online:2025-02-20 Published:2025-02-21

RichHTML

11

PDF

54

摘要：

【目的】随着生成对抗网络和扩散模型生成模型的快速发展，AI生成图像的质量不断提高，人类肉眼难以将其与真实图像区分。这些技术已经商业产品化，用户可通过软件产品一键式操作以实现文本到图像生成，产生了一定的商业价值，但也给司法鉴定带来了挑战，图像能否直接作为证据使用必将是法庭科学的重要研究课题。因此，如何有效地检测AI生成图像成为一个亟待解决的重要问题。【方法】现有的AI生成图像检测方法主要集中在检测单一生成模型的图像上。然而，一对一检测方法在面对未见过的生成模型时泛化能力较差。本文提出一种基于集成学习的Stacking策略集成多种一对一检测方法的对抗检测框架，通过使用随机森林模型结合为每个AI生成图像软件特化训练的单一检测器的输出，以替代直接泛化。【结果】实验结果表明，该框架在GenImage数据集上达到了98.36%的总体准确率，在本文设计的人工数据集上表现出较强的鲁棒性。单一检测器输出的分数被保留，为后续的归因工作提供可能。【结论】对抗检测框架具有可观的应用前景，作为一个可以灵活整合和更新各种检测技术的平台，为生成图像检测和归因研究提供一个更全和有效的解决方案。

关键词: AI生成图像检测, 集成学习, 图像检测框架

Abstract:

[Objective] With the rapid development of generative adversarial networks (GANs) and the generative diffusion models, the quality of AI-generated images has been continuously improved reaching to the point where it is challenging for the human eye to distinguish the AI-generated images from real images. This technology has been commercialized, allowing users to generate images from text with one-click software products, creating certain commercial value. However, they also pose challenges to forensic identification. Using images as a direct evidence is undoubtedly an important research topic in forensic science. Therefore, detecting AI-generated images has become a critical issue that needs to be addressed. [Methods] Existing methods for detecting AI-generated images mainly focus on detecting images from a single generative model. However, the one-to-one detection method is poor in generalization capabilities when facing unseen generative models. This paper proposes a Stacking-based ensemble learning strategy that integrates various one-to-one detection methods into an adversarial detection framework. It uses a random forest model to combine the output of individual detectors specifically trained for different AI-generated image software, instead of direct generalization. [Results] Experimental results show that the framework achieved an overall accuracy of 98.36% on the GenImage dataset and demonstrated strong robustness on the artificial dataset designed in this study. The scores from the single detector outputs are retained, providing possibilities for subsequent attribution work. [Conclusions] The adversarial detection framework is promising to be used as a platform that can flexibly integrate and update various detection technologies, providing a more comprehensive and effective solution for AI-generated image detection and attribution research..

Key words: AI-generated images detection, ensemble learning, image detection framework