An Approach of Fake Videos Detection Based on Haar Cascades and Convolutional Neural Network

Because deep fakes might skew our impression of the truth, we need to come up with a method for better spotting them. This paper proposes a new forensic technique to detect manipulated facial images from videos. It is based on CNNs architecture that can distinguish possible face manipulations in the “real-and-fake-face-detection” dataset offered by Kaggle. The results obtained highlight comparable performances with the state-of-the-art methods. It showed an accuracy of approximately 99 % for this binary classification into fake or real faces. Then to validate this model we added a human face detection technique using the Haar Cascade method to this model in order to detect the manipulated videos from Deep Fake Detection Challenge (DFDC) dataset and we achieve an accuracy of 91 correct predictions out of 100 total videos.