Comparative analysis of deepfake detection methods

Most security systems use biometric authentication pipelines to verify their users. Since one of the foundational aspects of biometric data is the human face, detecting deepfakes is essential to maintain the authenticity of identity and securing such systems. Deep learning models trained for detection of artificial faces commonly utilize transfer learning techniques where a model pre-trained on one task is fine-tuned for a new, related task. This situation arises because deepfake detection models have difficulty in progressing further and hitting a plateau due to the lack of domain-specific knowledge in their pre-trained dataset. In this study, we propose that pretraining state-of-the-art models using a specialized face dataset, namely Glint360K, enhances their performance significantly compared to base models pre-trained using the ImageNet dataset. While introducing a domain-specific dataset for models, we also aim to improve the models' ability of generalization by incorporating various regularization techniques. In particular, label smoothing was applied to models while training to reduce the effect of initial identities in the pre-training face dataset. Last but not least, our research presents a novel dataset with 4,252 real and 2,603 fake face videos created to mimic real-life situations with diverse lighting, device settings, and resolution differences. We utilized both this dataset and commonly used datasets to assess our proposed approach and state-of-the-art models, highlighting its strengths and weaknesses. The qualitative results show that our approach successfully force models to focus on face-related features. On the other hand, quantitative results indicate that models pre-trained on Glint360K with applied label smoothing outperform those pre-trained with ImageNet weights by 1.5% on average.