DETECTION OF DEEPFAKE IMAGES IN ELECTRONIC COMMUNICATION SYSTEMS USING NEURAL NETWORKS

Abstract

The article investigates the critically important problem of detecting generated visual content (deepfakes) in modern electronic communication systems. With the rapid development of Generative Adversarial Networks (GANs) and diffusion models, classical spatial detectors quickly overfit and lose their effectiveness when new generation algorithms emerge, which poses serious threats to information security and trust in digital communication channels. The aim of this work is to develop and comprehensively analyze modified neural network architectures in which, due to the synergistic combination of spatial and frequency attention mechanisms, a significant increase in the robustness and efficiency of detecting generated images is achieved. During the research, six unique hybrid architectures based on the deep convolutional network ResNet-50 were proposed and programmatically implemented. Classification efficiency was significantly improved through the targeted integration of multi-scale spatial attention (MSA), convolutional block attention module (CBAM), and frequency attention based on the Fast Fourier Transform (FFT), as well as the use of Transformer modules (Transformer Decoder) to analyze global structural relationships. Furthermore, a final ensemble model was developed, which effectively combines the predictions of specialized hybrids using the soft voting method, minimizing the false positive rates of individual classifiers. Experimental testing of the models was conducted on comprehensive datasets (DFFD, FaceForensics++, HiDF) and an independent balanced external dataset (Deepfake vs Real 60K) to objectively verify the capability for cross-domain generalization. According to the test results on the validation set, the developed ensemble demonstrated the highest efficiency: the overall accuracy reached 99.20%, and the F1-score was 0.9910. Testing on external data showed a decrease in accuracy to 74.7%, which confirmed the hypothesis about the difficulty of generalizing the features of new generators. By constructing Grad-CAM feature activation maps, the decision-making principles of the models were visualized in detail, and the critical importance of the frequency branch (ResNetSECBAMFFT architecture) for localizing spectral anomalies in cases of high-quality forgeries, which are completely ignored by classical spatial detectors, was mathematically proven.

Keywords: deepfake, convolutional neural networks, ResNet-50, attention mechanisms, Fourier transform, ensemble learning, Grad-CAM, cybersecurity.