Abstract: The deep learning-based facial forgery detection is commonly approached as a binary classification problem.The accuracy of model training results is not only affected by the quality and quantity of training data,but also related to training strategy and network architecture design..In this paper,we propose a new method based on key frames and spatial-temporal features.Firstly,the weighted optical flow energy analysis is used to detect the key frames in a video.Then,the optical flow and LBP features of the key frames are fused to form feature maps with spatial and temporal characteristics.After data augmentation,the feature maps are fed into the CNN model for training.Evaluations conducted on the FaceForensics++ and Celeb-df datasets de-monstrate that the proposed method achieves superior or comparable detection accuracy.Experimental results on cross-datasets show that the proposed method,utilizing the Efficientnet-V2 structure,achieves the best performance on the FaceForensics++ database with the accuracy of 90.1%.Furthermore,the overall performance of the XceptionNet structure surpasses that of other methods,achieving the accuracy over 80%,thus demonstrating superior generalization performance of the proposed method.

Key words: Optical flow, Key frames, LBP texture, CNN model