Comparative Evaluation of CNN and Transformer Architectures in Deepfake Detection Systems

Abstract

The increase in generation of deepfake media which is also known as synthetic media has created ample challenges in verifying authenticity of digital content. In response, researchers have started working in this area for developing robust detection mechanisms. This paper offers a comprehensive comparative evaluation of Convolutional Neural Networks (CNNs) and Transformer-based architectures for deepfake detection. After taking insights from over 20 recent articles and papers, this analytical study examines detection accuracy, generalization ability, computational efficiency, and robustness of deepfake detection architectures. Results reveal that the CNNs effectively capture local features; Transformer models outperform them in modelling global dependencies, achieving superior detection accuracy across different datasets.