Benchmarking Vision Transformers Against Traditional Architectures: a Systematic Analysis for Real-Time Deepfake Detection

The rise of deepfake technology is a serious threat to the trustworthiness of digital media. Using advanced deep learning techniques, deepfakes can produce highly realistic fake videos, images, and audios, posing challenges for detection. This paper presents a real-time deepfake detection system that uses Vision Transformers (ViTs) to address the limitations of existing methods. Unlike traditional approaches that struggle with generalization and computational efficiency, the proposed system effectively captures both spatial and temporal inconsistencies in video-based deepfakes. Through comprehensive evaluation, this research highlights improvements in accuracy, robustness, and real-time applicability, offering a practical solution to counter the growing deepfake menace. Our proposed model achieved an accuracy of 94 % with an inference time of 26.7 ms, allowing the model and integrated pipeline to detect deepfake in real time. This model captures long-range dependencies and global context, which allows it to mitigate the effect of advancements in Deepfake techniques.