Abstract:
As deep-fake image synthesis tools become more powerful and available, there is a growing need to develop methods for detecting generated content. The main goal of the work is to test the application of bispectral analysis as a tool for detecting images generated by artificial intelligence (AI). It is shown that higher-order spectral correlations detected by spectral analysis are less present in natural images compared to the images generated using generative-adversarial neural networks GAN (generative adversarial network). These correlations are probably the result of fundamental properties of the image generation process. The clustering procedure has shown encouraging results: it determines the generated images with an accuracy of 80%.