Multi-attention Based Face Forgery Detection
Xiang Han, Yongfeng Qi, Liqiang Zhuang, Liang Hu, Shengcong Wen · 2023
Problems such as low accuracy and weak generalisation performance are common in existing deep face forgery detection algorithms. In order to effectively detect the features of forged faces, a deep neural network detection method based on multi-attention mechanism is proposed. Firstly, ResNet50 convolutional neural network is used to extract preliminary forged features from the input face image; secondly, channel attention in hybrid attention is used to capture the abnormal features of the forged face, and combined with spatial attention to focus on the location of the abnormal features, fully learn the semantic information of the abnormal part of the forged face in the shallow layer, and then get the shallow feature maps by sequentially connecting the two; and then, the shallow feature maps are inputted into the Then the shallow feature map is input into the dual-layer attention to get the deep feature map; finally, the deep features are input into the classification module for accurate classification. The detection accuracy of the proposed method under two different compression rates of FaceForensics+ + dataset reaches 97.74% and 91.46% respectively, while the area under the receiver operating characteristic curve reaches 98.97% and 95.87% respectively, which is better than methods such as XceptionNet and Efficientnet. The experimental results show that the proposed method has good accuracy and strong generalisation performance.