哈尔滨商业大学学报（自然科学版）2023，Vol.39Issue(6)：661-667,7.
基于局部和全局特征的深度伪造检测方法
Deepfakes detection methods based on local and global features
杨新露 1程健 2张凯1
作者信息
- 1. 安徽理工大学 计算机科学与工程学院,安徽 淮南 232001
- 2. 安徽理工大学 人工智能学院,安徽 淮南 232001
- 折叠
摘要
Abstract
In recent years,deepfakes have generated high-quality fake images or videos.They were used for entertainment but also posed a serious threat to the security of others.Al-though the existing detection methods achieved high accuracy on some datasets,their effec-tiveness was greatly reduced when tested across datasets.To solve this problem,a deepfake detection method based on a combination of local and global features was proposed.This method extracted global information from image blocks of different sizes through a multi-scale Transformer module.The EfficientNet network was used to extract local features of the image in combination with the attention mechanism.Finally,the deepfake face was classified ac-cording to the extracted local features and global features.The experimental results showed that the detection accuracy on the FaceForensics + + dataset reached 96.78%.The results of testing on the Celeb-DF and DFDC datasets reached 77.3%and 73.32%,respectively.This method extracted more local and global artifacts in fake faces,which had certain gener-alizations between different datasets.关键词
深度伪造检测/局部特征/全局特征/Transformer/注意力机制Key words
deepfakes detection/local features/global features/transformer/attention mechanism分类
信息技术与安全科学引用本文复制引用
杨新露,程健,张凯..基于局部和全局特征的深度伪造检测方法[J].哈尔滨商业大学学报（自然科学版）,2023,39(6):661-667,7.
