HYBRID CNN-KAN ARCHITECTURE FOR DETECTING GENERATED IMAGES УДК 004.056.57 : 004.852

Abstract

The article discusses the development of a hybrid CNN-KAN neural networkarchitecture designed for detecting generatedimages. The problem of detecting synthetic visualcontent is becoming increasingly importantdue to the rapid growth of generative AItechnologies. The paper describes the principlesof convolutional neural networks (CNNs) andKolmogorov-Arnold networks (KANs), andprovides a rationale for combining them intoa single hybrid model. It has been experimentallyshown that CNN-KAN performs better thantraditional CNN-MLP architectures in termsof Accuracy, Precision, Recall, F1-score, andROC-AUC metrics. The proposed architecturecombines the ability of CNN to extract localfeatures with the ability of KAN to generalizeglobal patterns, resulting in high classificationaccuracy. The research findings demonstratethe effectiveness of the hybrid approach and itspotential for use in fake image detection tasks.