DETECTION OF SPOOFING ATTACKS BASED ON CONVOLUTIONAL NEURAL NETWORKS AND COLOR SPACES УДК 004.056.57

Abstract

The article describes the development of a method for detecting spoofingattacks based on convolutional neural networksand color spaces.Two convolutional neural network (CNN)architectures were developed, using differentcolor spaces. The first architecture has oneinput layer for one color space. The secondarchitecture has two parallel input layers for twocolor spaces. The following models were trained:RGB, CMYK, YCbCr, HSV, RGB+YCbCr,RGB+HSV, RGB+CMYK, YCbCr+CMYK,YCbCr+HSV, HSV+CMYK. The models weretrained on two datasets: OpenForensics Datasetand Dataset Nvidia & StyleGAN. It was foundthat all developed models show fairly high resultsin detecting generated images. The results of thework can be applied to detect generated imagesusing convolutional neural networks and colorspaces.