WPCA: Wavelet Packets with Channel Attention for Detecting Face Manipulation.

Malicious applications of face manipulation technologies pose a great threat to social stability, which promotes the development of face forgery detection. Previous efforts for manipulation detection mainly focus on the features in the spatial domain, while the information in the frequency domain has not been effectively mined. Moreover, many existing forgery techniques are based on GANs, and their forgery artifacts are easier to be captured in the frequency domain. Hence, we propose a lightweight framework called Wavelet Packets with Channel Attention (WPCA) to tackle this issue. Our model employs the wavelet packets to obtain the decomposed frequency representations, and introduces the channel attention mechanism to focus on the discriminative feature blocks. We further design a specific refined network for manipulation detection by taking the weighted features as input. Extensive experiments have demonstrated that our proposed model achieves desirable detection performance and converges faster with fewer model parameters.