Enhancing Deepfake Detection Through Dynamics of Facial Expressions
Vishal Kumar Sharma, Seema Rawat · 2025
The rapid evolution of deepfake technology poses significant threats to digital security, media integrity, and public trust, necessitating the development of robust detection frame-works. Traditional deepfake detection methods primarily rely on pixel inconsistencies, frequency-domain analysis, or handcrafted features, but these approaches are increasingly vulnerable to advanced generative models that produce high-fidelity manipulations. In this study, we introduce MADDM, a Masked Autoencoder-based deepfake detection model that leverages facial expression dynamics to identify inconsistencies in muscle coordination—an aspect that remains challenging for deepfake generators to replicate accurately. Our model is trained in a self-supervised manner, first learning natural facial expressions from real datasets and then detecting anomalies in synthetic videos by reconstructing masked facial regions. Evaluations on Celeb-DF, DFDC, and FaceForensics++ datasets demonstrate that MADDM significantly outperforms existing detection methods, achieving an average accuracy of 81.1%, with state-of-the-art performance on Celeb-DF (86.3%). Further analysis through intra-dataset and cross-dataset testing confirms the model’s superior generalization capabilities. The results highlight the potential of expression-based deepfake detection as a powerful and scalable solution for digital forensics and misinformation control. Future research should explore real-time implementation, transformer-based optimizations, and adversarial training strategies to enhance detection efficiency against evolving deepfake techniques.