3sh-Vss Network with Statistical Analysis for Deepfake Video Detectio

Abstract

In this paper, we first select the features with the largest differences between real and fake videos, i.e., the fake features on YCbCr color space, through statistical analysis. Then, we introduce the Mamba into the field of deepfake identification for the first time, and design the 3SH-VSS deepfake detection model, so that global features and local features are considered at the same time. The 3SH branch aims to increase the model receptive field by depthwise separable convolution streams with different convolution kernel sizes, while increasing the interaction between different streams through channel shuffling for better extraction of local features. The VSS branch captures global features by scanning a series of features in four different directions through two-dimensional selective scanning (SS2D). Extensive experiments show that the method proposed in this paper achieves state-of-the art performance on different benchmark datasets and outperforms the state-of-the-art for compressed deepfake video detection while maintaining generalization.

Keywords: Deepfake, face forgery detection, VMamba, Channel shuffle