深度伪造人脸图像检测方法研究进展：从强监督到无监督学习

Advances in Deepfake Facial Image Detection Methods: From Fully Supervised to Unsupervised Learning

摘要: 随着生成式人工智能快速发展，深度伪造人脸图像的真实感不断提升，对金融安全、司法取证和社会信任构成威胁，深度伪造人脸图像检测因此成为数字内容安全领域的重要研究方向。本文首先分析自编码器、生成对抗网络、语义可控生成和扩散模型等生成技术演进对检测线索变化的影响；其次按照训练过程中对标注信息的依赖程度，将现有检测方法归纳为强监督检测、弱监督与半监督检测、无监督检测三类，并总结其技术特点与局限。综述表明，相关研究正由显性伪影识别转向隐性特征建模，由闭集二分类转向跨域泛化与开放场景适应。未来需重点提升隐性特征挖掘、低标注学习、跨域泛化、复杂扰动鲁棒性和可解释评测能力，并关注视觉基础模型、多模态一致性检测与主动防御技术的发展。

Abstract: With the rapid development of generative artificial intelligence, deepfake facial images have become increasingly realistic, posing threats to financial security, judicial forensics, and social trust. Deepfake facial image detection has therefore become an important research direction in digital content security. This paper first analyzes how the evolution of generative techniques, including autoencoders, generative adversarial networks, semantically controllable generation, and diffusion models, has changed detectable forgery traces. Then, according to the dependence on annotation information during training, existing detection methods are categorized into three groups: fully supervised detection, weakly supervised and semi-supervised detection, and unsupervised detection, with their technical characteristics and limitations summarized. This review shows that related research is shifting from explicit artifact recognition to implicit feature modeling, and from closed-set binary classification to cross-domain generalization and open-scenario adaptation. Future research should focus on improving implicit feature mining, low-annotation learning, cross-domain generalization, robustness against complex perturbations, and interpretable evaluation, while also paying attention to the development of vision foundation models, multimodal consistency detection, and active defense techniques.