Deep learning methods have undergone significant development in recent years. In particular, Generative Adversarial Networks (GAN), are now able to generate natural-looking images and create fake identities in images or in videos. For obvious reasons, identifying spoofed videos is essential to avoid manipulation of information. In this thesis we want to propose methods for detecting Deepfakes based on the traces left in the images by generative networks and video encoding and compression methods, and making use of specific neural network approaches integrating attention mechanisms.