Программное средство, определяющее фейковый видеоконтент с помощью технологии deepfake алгоритма GAN

Abstract

In the modern world, one of the main and urgent problems is false content: news, videos, photos, etc. Early on in the development of Deepfake technology, it was used by amateur users to generate multimedia content by matching human facial expressions and phrases, usually owned by recognizable individuals, to create fake media that looked genuine. But the situation is changing, and Deepfake technology is being used not for compromising, but for campaigning and attracting political supporters. The purpose of the study: Software implementation of the video content recognition algorithm, synthesized using the Deepfake technology of the GAN algorithm, with acceptable accuracy. In the work, a software implementation was proposed that analyzes the video and makes a decision about the authenticity of this one. The main architectures of the GAN algorithm are presented, as well as the opportunities and threats of using deepfake technology. An analysis of the features of the Xception and ResNeXt models trained using neural networks was carried out. Methods: For the system to work, it is necessary to select suitable neural networks based on performance results, which can be ResNeXt, XceptionNet or any other neural network. As part of this work, ResNeXt and XceptionNet will be considered and used in the software implementation, as well as BlazeFace is a pre-trained human face recognition model used to recognize faces in extracted images. Results: The function input is the path to the video (in the file system). The sample is frame-by-frame checked for the presence of a face in each individual frame, if the recognition was successful, the data is added to the list. Optionally, you can leave a fixed number of samples with the best quality among those presented.