Advancing Deepfake Detection Through Head Pose Estimation on New-Generation Datasets

The increasing prevalence of Artificial Intelligence (AI) generated content, particularly deepfake videos, has raised serious concerns in academia and society. This study explores the use of Head Pose Estimation (HPE) as a discriminative feature for deepfake detection, using a distance-based classification approach via K-Nearest Neighbours (KNN) combined with Dynamic Time Warping (DTW). Three HPE methods - Feature Selective Attention Network (FSA-Net), SynergyNet and Web-Shaped Model (WSM) - were tested on three widely used public datasets: WildDeepfake, Celeb-DF and DeeperForensics-1.0. The results show that the WSM method offers superior performance compared to the other approaches, showing a good balance between the Real and Fake classes, particularly on complex datasets such as DeeperForensics-1.0, demonstrating its stability compared to previous results in literature on this method. The results obtained open up various perspectives and future directions for tackling the problem of deepfake detection by exploiting HPE-based approaches, which are notable for their speed and high reliability.