Development of a Deepfake Detection Method: Application of Frequency Analysis and Reduction of the Image Color Space to Improve Classification Accuracy

With the development of generative and diffusion models of neural networks, it has become possible to create high-quality realistic images that are visually indistinguishable from real photos. This expands the horizons of using technology to synthesize data and media content, however, with the help of this technology, it becomes possible to generate or replace the faces of real people using deepfake methods. Although modern deepfake recognition methods show high results on synthetic data, they suffer from the problem of false positives when working with processed images of real and synthetic faces. Moreover, resizing the image according to the input layer of classifiers based on convolutional neural networks can lead to additional distortions, which can affect the classification process. In this paper, we propose a method based on the reduction of the color palette and spectral analysis of images to improve the accuracy of the classification of deepfakes. The proposed method allowed to increase the accuracy of recognition of modified deepfakes to 99.4%.