Exploiting Correlation Between Facial Action Units for Detecting Deepfake Videos
Quoc Hoan Vu, Priyanka Singh · 2024
The potential misuse of deepfakes has led to an increase in research into deepfake detection methods. These methods employ machine learning, deep learning, and other intricate techniques to spot subtle inconsistencies in deepfake videos. However, as deepfake technology advances, the current research exposes gaps, such as inadequate performance on real-world data and susceptibility to adversarial attacks. Among various detection methodologies, the use of identity-based features like facial action units shows potential in identifying synthetic content. Given that correlations between facial features may be distinctive, they could play a crucial role in differentiating real and artificial imagery. In this study, we propose a method of high accuracy to distinguish genuine videos from deepfakes by exploiting the correlation of facial action units. Our main contributions are twofold: (1) merging identity-based techniques with traditional machine learning for deepfake detection, and (2) evaluating this approach's robustness against compression.