Academic Editor: Youssef EL FOUTAYENI
Received |
Accepted |
Published |
29 January 2020 |
13 February 2020 |
10 March 2020 |
Abstract: A deepfake is a type of montage which exploits an artificial intelligence technology to obtain fake videos where the face of a person is modified and replaced by that of another person and modify the words of a person with a suitable facial expression [1]. To be able to identify videos with facial or vocal manipulations we have built a deep learning model (convolutional neural network (CNN)) with a multitude of videos of two classes (video fake and truthful video) by interpreting the images of each video one by one. 'learning. Next, our model will analyze each of the videos and draw their own result. The goal is that at the end, once the learning phase is over, if we give our model a video it will be able to deduce whether it is fake or not (with a certain probability). The working principle of our proposal is to analyze the variations in the blood flow on a face. With each heartbeat, our face becomes a little redder. It is imperceptible to the eye, so if a face has been superimposed on another video to create a deepfake, the false face will have no heartbeat and our model will report the video as faked. For the experiments we used the dataset created by facebook made up of 100,000 videos with actors and for realistic scenarios. these videos were manipulated and annotated by AI. The results are satisfactory compare to other techniques that exist in the literature. To conclude, we are sure that in the future deepfakes will be a rapidly evolving challenge, just like spam, phishing and other threats.
