Semantic and Frequency Representation Mining for Face Manipulation Detection

Jihao Cao,Jinsheng Deng,Xiaoqing Yin,Zhichao Zhang,Hui Chen

DOI: https://doi.org/10.1007/978-3-031-44213-1_10

2023-01-01

Abstract:Face manipulation technologies pose a great threat to the current digital media. Although previous methods have achieved excellent detection performance, they tend to focus on specific artifacts and lead to overfitting. Erasing-based augmentations can alleviate this issue, but they still suffer from high randomness and fixed shapes. Therefore, we propose a novel face masking method named Landmarks Based Erasing (LBE), which exploits the geometric information of the face and forgery attention map to perform erasure, thereby forcing the network to mine discriminative features from other face regions. Furthermore, Wavelet Packet with Attention (WPA) mechanism module is designed to extract multi-level frequency features, providing a complementary perspective to LBE module. Finally, we employ a score fusion strategy to fuse two types of complementary feature information for forgery detection. Extensive experiments on three large public datasets demonstrate that our proposed method achieves state-of-the-art detection performance and exhibits good generalization ability.