An Ensemble Approach to Facial Deepfake Detection Using Self-Supervised Features

Substantial efforts have been dedicated to developing methods for detecting deepfake content, especially with the creation of large and diverse datasets with both higher image quality and demographic features. In this scenario, CNN-based approaches showed good initial success, later improved by their combination with Vision Transformers. More recently, Foundation Models (FMs) have emerged, improving performance across many visual tasks, including deepfake detection, and combining self-supervised features generated by FMs with CNN-based classifiers has resulted in significant performance gains. However, taking advantage of multiple maps of self-supervised features is not as straightforward as just adding more channels to the classifier. Therefore, this work explores ensemble techniques to effectively utilize these diverse self-supervised feature maps for realistic facial deepfake detection. Our experiments indicate that combining the output results of different classifiers, each one utilizing a single map of self-supervised features, leads to significant performance improvements, and several committee approaches consistently outperform individual classifiers, demonstrating the potential of these methods in enhancing deepfake detection accuracy.