Learning to detect Deepfakes: benchmarks and algorithms

Abstract

The capabilities of deep-learning tools have led to the emergence of the so-called Deepfakes. These are a type of videos involving a person whose face has been artificially forged in one way or another. These videos poses a serious threat to information veracity and integrity in social media. Therefore, it makes sense that companies and institutions have a tool available to identify such type of resources in order to take them down from the Internet. As generation methods have become more and more sophisticated, building models for the detection of these videos is an increasingly popular area of research. The task is not easy and requires bringing together several modules as well as taking into consideration distinct factors. In this work, we present a survey of the state-of-the-art of current generation and detection methods. Simultaneously, we analyse the results obtained with different models by formulating the problem as a binary classification task at a frame level. These results allow the comparison of some Convolutional Neural Networks architectures as well as several data augmentation policies. To do so, we have run our models in two different benchmark datasets: one that is originally from the academia and the another one derived from the industry. Nonetheless, despite the effort put by researchers on detection methods, more work has to be done in order to achieve feasible solutions. For example, so far end-toend trainable models have not yet been accomplished and there exists a generalization problem in detection models.