Hybrid two-level method for automatic detection of face substitution in an image

Purpose of research. The development of a hybrid, two-level method to enhance both the accuracy and robustness of detecting operator face spoofing in images, which is a pressing issue given the constant growth and sophistication of threats from deepfake technologies. Methods. A novel architecture is proposed, combining the EfficientNet convolutional neural network for deep pattern extraction with an ensemble of four classifiers. These classifiers specifically analyze distinct feature groups: expertbased, textural, statistical, and those based on facial landmark coordinates, enabling the detection of specific synthesis artifacts. For training and testing, an extensive and representative dataset of 34,000 images was compiled, including deepfakes generated by several modern tools as well as public datasets. Results. The high efficacy of the proposed method was experimentally confirmed: accuracy reached 0.921 and the F1-score was 0.914. These metrics significantly surpass the performance of any of the individual models used separately, demonstrating a pronounced and practically significant synergistic effect from their combination. Conclusion. This work demonstrates that the synergy between deep learning and classical feature-based models allows for the creation of a genuinely more reliable and precise detector. The proposed method improves overall accuracy and enhances system robustness by effectively compensating for the individual weaknesses of separate classifiers. This validates the hypothesis that combining a neural network's ability to extract complex, implicit patterns with feature-based models' capacity to analyze specific, predefined artifacts (such as geometric distortions) leads to a more powerful and resilient detector.