An Ensemble Deep Learning Framework for Deepfake Video Detection Using ResNet Models

Deepfake technology has raised growing concerns due to its potential to enable misinformation, identity fraud, and security threats through manipulated digital content. As these risks continue to increase, the development of reliable deepfake detection methods has become essential. In this study, a spatial feature-based ensemble deep learning framework is proposed to improve detection accuracy and classification stability. Facial regions were first extracted from video frames using the MTCNN algorithm to focus the analysis on relevant facial areas. The proposed framework focuses on spatial facial representations extracted from individual video frames rather than explicit temporal modeling. Subsequently, ResNet50 and ResNet101 were employed as the main spatial feature extractors to obtain deep facial representations at different abstraction levels. ResNet152 was also evaluated experimentally to examine the effect of a deeper feature extraction branch; however, it was not included in the final framework because it increased runtime without improving video-level classification performance. The extracted feature vectors were classified using independent multilayer perceptron (MLP) classifiers, and the final prediction was obtained using a two-model weighted soft voting ensemble. The proposed method was evaluated on a balanced subset of the Celeb-DF dataset using a video-level train-test split to reduce data leakage between training and test sets. Experimental results showed that the final ResNet50–ResNet101 weighted soft voting ensemble achieved 94.17% video-level accuracy, 0.986 AUC, 0.9440 F1-score, and 0.8864 MCC. The two-model ensemble achieved performance comparable to the three-model configuration while providing a more favorable runtime profile. These findings suggest that combining complementary spatial facial representations through weighted soft voting provides a reliable and computationally efficient approach for deepfake video detection.