Development of Neural Networks for Deepfake Recognition

Abstract

Nowadays, the creation of deepfakes for various activities is widespread: sometimes for the sake of entertainment, and sometimes for malicious purposes. In the second case, such deepfakes can potentially harm a person. Neural networks trained to recognize deepfakes can become an actual tool for combating malicious fakes.This article considers neural network architectures which can be used to solve the problem of recognizing photo and video deepfakes, describes datasets that can help in training neural networks for the task, and determines what preprocessing is needed for different datasets. Particular attention is paid to possible options for combining the results of several trained models.