MSFF-Net: A Deepfake Detection Network Based on Multi-Space Feature Fusion Technique

Abstract

The history of digital manipulation has entered a new chapter with the development of deepfake technology, opening the door to realistic yet misleading visual content creation. Deepfake technology's rapid rise has raised worries about potential misuse because it may produce fake or misleading information with incredible realism. Because deepfakes made it possible for actors' faces to be perfectly altered in movies, the entertainment industry was the first to adopt them. Since then, however, they have evolved and been utilized in a variety of settings, including social networking sites, political issues, and journalism. Feature extraction from digital photographs had a major impact in the field of detecting deepfakes, where separating actual information from manipulated media is crucial. Many techniques for detecting deepfakes based on fabricated features have been developed recently. Textural features are a frequently used type of forged feature. However, most of the current detection methods relies on single space features, ignoring other potentially informative features. To overcome this drawback, this study presents a novel deepfake detection network named MSFF-Net (Multi-Space Feature Fusion Network). This network integrates Histogram of Oriented Gradients features with the deep features extracted using ResNet50 model. HOG descriptors analyze local intensity gradients, which helps in understanding the texture and shape of the image, while the deep features extracted by the ResNet50 model capture high-level semantic information. By combining these diverse feature sets, one can create a richer representation of the data that encompasses both low and high-level characteristics, thereby enhancing the robustness and effectiveness of deepfake detection. The experiments are performed on the publicly available datasets 140K real and fake faces dataset, DFDC, and Celeb-df. MSFF-Net showed better performance than the other SOTA models, thereby improving the ability to detect deepfakes.