Lightweight detection method for deepfake face video

This paper proposes a lightweight fake facial video classification detection model to address the issues of insufficient accuracy, poor robustness, and excessive parameters in existing forgery face video detection models. The model utilizes an improved version of EfficientNetV2-S as the backbone network for feature extraction. The MTCNN algorithm is employed to extract the facial region and perform alignment, while data augmentation techniques are applied to enhance the model's robustness. The pre-processed data is then fed into the backbone network for classification detection. Furthermore, in order to improve the performance of the network model, reduce the influence of network parameters, and decrease model complexity, the structure of the backbone network is streamlined and the composition of the network is adjusted. Additionally, the CBAM attention module is introduced to enable the shallow model to capture more facial details. Experimental results demonstrate that compared to the original EfficientNetV2-S model, the proposed model achieves a 50% reduction in parameter count. Moreover, it achieves accuracy improvements of 1.1%, 2.8%, and 3.1% on the FaceForensice++, Celeb-DF-v1, and DFDC datasets, respectively. When compared with the Xception, CapsuleNet, and DefakeHop models in the field, varying degrees of improvement are also observed.