Toward Efficient Fake Frame Detection in Video Using Deep Learning

Abstract

Background: Deepfake technology is a major social concern due to the rapid development of Artificial Intelligence (AI), especially in machine learning (ML) and deep learning (DL). Deepfakes are artificially modified videos and images that effectively change a person’s facial features or expressions to misrepresent reality. These videos and images, often unnoticed by casual observers, present significant ethical, political, and social implications, as they can spread misinformation, damage reputations, and influence public perception. This study is an attempt to detect artificially modified videos by analyzing each frame using DL. We use ResNet-50 architecture, a well-known Convolutional Neural Network (CNN) model, to identify tampered videos. Methods: The system is trained using the Celebrity Deep Fake dataset, which includes numerous original and fake video samples. The model assesses whether each frame is original or tampered with after the videos have been split into frames. The system is tested and evaluated using standard metrics, including accuracy, precision, recall, and F1-score. Results: The model achieved 82.33% accuracy, 76.70% precision, 89.15% recall, and an F1-score of 82.46%. These results indicate that deepfake videos were correctly detected and that the model was efficient at identifying most real deepfake instances. In addition, the F1-score of 82.46% is further evidence of the model’s stability, as it ensures both high accuracy and coherence across numerous cases. Conclusions: The findings indicate that the proposed model works effectively compared to existing deepfake methods. In the future, we intend to use a richer dataset with more resources, which may further enhance the accuracy of the model.