This article examines the problem of deepfake image recognition using a neural network model. It provides an overview of modern research on deepfakes in images, identifying the advantages and disadvantages of existing recognition methods. Existing datasets of real images and deepfakes are analyzed, and the choice of the most suitable one for training, validating, and testing the neural network model is substantiated. As a result of the dataset analysis, the ArtiFact (Real and Fake Image Dataset) dataset was selected as the initial set. It contains over 2 million images, 19 image generators, and 11 sources of real images. Given the limited computing resources of the environment, it was decided not to use the entire found dataset, but to select the necessary images to create our own datasets: final (training, validation, test sets), controlA, and controlB. Neural network architectures for use in the deepfake image recognition problem are analyzed. The Xception model is used to recognize deepfake images. The input image size for the model is set by default - 299×299. The Google Kolab cloud environment, a local code execution environment (Jupyter Notebook), and the Python programming language were used to train the neural network model. A neural network model is trained, and the optimal training process and hyperparameters are selected. Two approaches to deepfake image recognition are compared: a modified Xception neural network model with a transformer head and an ensemble of four models (Xception, Efficientnet-B4, ConvNeXt, and Swin Transformer). The analysis revealed that the CNN+tr.head model demonstrates high metric results on the test set (F1 = 0.9105), but is sensitive to new types of generators, resulting in a drop in the F1 metric on the validation set to 0.7414. The ensemble approach, in contrast, provides higher robustness when recognizing images generated by previously unknown models (F1 metric on the validation set is 0.8082).
DEEPFAKE, ARTIFICIAL INTELLIGENCE, NEURAL NETWORK MODEL, GENERATED IMAGE, DEEPFAKE RECOGNITION, NEURAL NETWORK
1. Deepfake. URL: https://ru.wikipedia.org/wiki/ Deepfake (accessed March 2, 2026).
2. What Are Deepfakes, What Are They Used For, and Why Are They Dangerous? URL: https://practicum.yandex.ru/blog/chto-takoe-deepfake-i-kak-zashchititsya/#chtotakoye-dipfeyk (accessed April 24, 2026).
3. Nerenz D.V., Philology: Scientific Research, 9, 96–111 (2025).
4. Yong Y., Zhihao Q., Ye Z., Russakovsky O., Yu W., Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 23850–23859 (2025). DOI: https://doi.org/10.1109/CVPR52734.2025.02221
5. Junyu S., Minghui L., Junguo Z., Zhifei Yu, Yipeng L., Shengshan H., Ziqi Z., Yechao Z., Wei W., Yinzhe X., Leo Yu Z, 39th Conference on Neural Information Processing Systems (NeurIPS), 1–13 (2025).
6. Pikul A.S., Information Technology Security, 31, 4, 116–127 (2024). DOI: https://doi.org/10.26583/bit.2024.4.08
7. Voronov S.A., Information and Security, 28, 1, 103–110 (2025). DOI: https://doi.org/10.1504/IJICS.2025.148456
8. Andriyanov A.M., Scientific and Technical Bulletin of the Volga Region, 1, 169–172 (2025).
9. Grishaev D.A., Bulletin of Penza State University, 1 (49), 28–31 (2025).
10. Garifullin N.B., A.A. Vavilov Youth School-Seminar on Control Problems in Technical Systems. 1, 13–15 (2024).
11. Mantulenko A.I., Mathematical Methods in Technology and Engineering, 5, 97–101 (2025).
12. Petrin D.A., Proceedings of the Institute of Engineering Physics, 1 (59), 56–60 (2021).
13. genimage dataset. URL: https://drive.google.com/drive/folders/1jGt10bwTbhEZuGXLyvrCuxOI0cBqQ1FS (accessed 03/09/2026).
14. Datasets: InfImagine / FakeImageDataset. URL: https://huggingface.co/datasets/InfImagine/FakeImageDataset (accessed March 9, 2026).
15. ArtiFact: Real and Fake Image Dataset. URL: https://elibrary.ru/item.asp?id=49175611 (accessed March 9, 2026).
16. Datasets: yanbax / CIFAKE_autotrain_compatible. URL: https://huggingface.co/datasets/yanbax/CIFAKE_autotrain_compatible (accessed March 9, 2026).
17. Synthbuster: Towards Detection of Diffusion Model Generated Images. URL: https://zenodo.org/records/10066460 (accessed 03/09/2026).
18. GitHub: opendatalab / LEGION. URL: https://github.com/opendatalab/LEGION (accessed March 9, 2026).
19. Gagarina A.I., Modern Pedagogy and Scientific Research in Higher Education Institutions: Proceedings of the All-Russian Scientific and Methodological Conference, 694–704 (2022).
20. Akobiya V.Z., Current Trends in Development and Prospects for the Implementation of Innovative Technologies in Mechanical Engineering, Education, and the Economy, 7, 1, 136–138 (2025).
21. Convolutional Neural Network, Part 1: Structure, Topology, Activation Functions, and Training Set // habr.com [website]. – URL: https://habr.com/ru/articles/348000/ (accessed: 04/24/2026).
22. Ivanov P.P., Digital Region: Experience, Competencies, Projects: Collection of Articles from the VII International Scientific and Practical Conference Dedicated to the 95th Anniversary of Bryansk State University of Engineering and Technology, 314–318 (2025).
23. Xception: A Compact Deep Neural Network. URL: https://habr.com/ru/articles/347564/ (accessed: April 24, 2026).
24. Using EfficientNet models for image classification. URL: https://habr.com/ru/companies/sberbank/articles/828842/ (accessed: April 24, 2026).
25. Overview – ConvNet for 2020. URL: https://habr.com/ru/companies/otus/articles/654279/ (accessed: 04/24/2026).
26. Overview of the Swin Transformer architecture. URL: https://habr.com/ru/articles/599057/ (accessed: 04/24/2026).
27. Ensemble methods in machine learning. URL: https://habr.com/ru/articles/571296/ (accessed: 04/24/2026).
