Data-Efficient Transformer Architectures for Image-Level Facial Forgery Detection: A Comparative Evaluation of ViT and DeiT

The rapid development of deepfake technologies has increased the demand for a credible and inter-pretable system for facial forgery detection. This study compares two transformer-based architec-tures—Vision Transformer (ViT) and Distilled Data-Efficient Image Transformer (DeiT)—for de-tecting real and manipulated facial images. The study aims to measure performance in terms of de-tection as well as interpretability and to address the weaknesses of traditional convolutional models. Data augmentation was applied, and a balanced dataset containing ۸,۰۰۰ real and fake images was constructed; both models were then fine-tuned under the same training environment. The explanatory ability of the models was incorporated using LIME. Experimental findings indicate that both models perform well, with DeiT being slightly more accurate at ۹۴.۶۲% than ViT at ۹۳.۶%, alongside faster convergence rates and less overfitting. Visualization of the focus on important facial areas confirms that the models reliably register synthetic artifacts. Although promising, generalization across dif-ferent datasets and enhancement of real-time performance remain challenges. Overall, the results validate transformer architectures—especially DeiT—as powerful and explainable deepfake detec-tion algorithms, valuable for ensuring safe and transparent digital media forensics.

کلیدواژه ها:

Vision Transformer ، Data-Efficient Image Transformer ، Face Forgery Detection ، Ex-plainable AI ، Transformer Models