A Hybrid Model of Deep Learning and Machine Learning Methods to Detect Deepfake Videos

Abstract

The deepfake videos were spread in the last few years and were created by different deepfake techniques (i.e., faceswap, face2face, etc.). These techniques have a terrible impact on society and would give anyone a chance to create videos with fake faces. The objective of this paper was to develop a model that detects deepfake videos to reduce their negative impact. Two hybrid models of machine learning and deep learning were proposed. The first model used a convolutional neural network (CNN) for feature extraction and different machine learning classifiers (support vector machine (SVM), k-nearest neighbor (KNN), decision tree (DT), random forest (RF), logistic regression, and naive bayes (NB)). In contrast, the second model used a transfer learning concept developed by the VGG16 (Visual Geometry Group) pre-trained model and the same machine learning classifiers as the first model. Both models were evaluated on the FaceForensics++ video dataset, which includes four different deepfake techniques (Deepfake, Faceswap, Face2Face, and Neuraltexture). The results showed good accuracy, which proved the effectiveness of the proposed models, which may be used as a detection deepfake application. While the first model can obtain the highest accuracy with the SVM classifier on the four deepfake techniques sequentially: 0.96, 0.87, 0.90, and 0.64. In contrast, the second model achieved the highest accuracy with the KNN classifier on Deepfake and Face Swap techniques: 0.95 and 0.91, and with SVM on Face2Face and Neural Texture techniques: 0.86 and 0.77.