부산대학교 도서관

The world is witnessing great developments daily in the field of graphics and computer vision. Now, it’s possible to create fake videos with very realistic faces. Thus, discrimination between original and fake videos has become a major challenge, which caused serious threats to both the individual and society. Usually, the traditional image forensic technicalities are not appropriate to classify videos because of data compression that damages it. Thus, this research focused on the use of hybrid deep learning models that based on convolutional neural networks (CNN) and recurrent neural networks (RNN) methods for fake video detection. The inceptionV3 model was used to extract facial features from the frames, then these features were used to train simpleRNN and Gated Recurrent Unit (GRU) models to classify video. Most deepfake detection works fails when tested on a new dataset, especially those that are real and close to reality. Therefore, the most realistic dataset which produced ‘in the wild’ was chosen in this research. The deepfake detection challenge (DFDC) dataset was used to evaluate the proposed models. Where these models achieved a high detection accuracy, 98.5% for SimpleRNN and 98.9% for GRU. Also, the models achieved 0.979 and 0.986 of AUC respectively.