基于时空特征和注意力机制的伪造检测方法

打开文本图片集

中图分类号：TP183 文献标志码：ADOI： 10. 13705/j. issn. 1671-6841. 2024045

文章编号：1671-6841（2025）05-0009-07

Abstract：A facial forgery detection method based on spatio-temporal features and atention mechanism was proposed to address the issues of redundant features and lack of discriminative power in forgery detection. The aim was to explore forgery clues of images in the temporal and spatial domains，using interframe difference information as a starting point for forgery detection. Firstly，a spatial attention module was utilized to direct the model's atention towards the facial regions more prone to forgery in real scenes. Secondly，a temporal attention module was employed to assign greater weights to frames with larger facial motion，which exhibited greater discriminative power.Experiments were conducted on the FaceForensics ++ dataset，and the results showed that the area under curve （AUC） ）index of the proposed method reached 89.04% and 98.81% on low-quality and high-quality videos，respectively. In addition， the experimental test results on the Celeb-DF dataset demonstrated that the proposed method had good generalization capability.

Key words： forgery detection; spatio-temporal feature; inter-frame difference； atention mechanism

0 引言

随着生成模型的发展，人脸伪造技术取得了显著的进展。（剩余12625字）