<![CDATA[Deepfake technology poses an increasingly serious threat to personal reputation and social trust, necessitating the development of accurate yet computationally efficient detection systems. Although large pre-trained vision models offer exceptional feature extraction capabilities, full fine-tuning them demands prohibitive computational resources and risks overfitting. This study investigates the application of Low-Rank Adaptation (LoRA) to enhance binary image classification accuracy for deepfake face detection, bridging the gap between parameter efficiency and high classification performance. We systematically integrate LoRA into two dominant architectural paradigms: the Vision Transformer (Swin-T) and ResNet-50. Computational evaluations are conducted on a 40K sub-dataset from the 140K Real and Fake Faces dataset, comparing LoRA against full fine-tuning baselines under identical environments. Experimental results demonstrate that Swin-T + LoRA achieves an outstanding test accuracy of 99.14% and an F1-score of 0.9913, outperforming its full fine-tuning baseline by 9.95 percentage points while training only 5.88% of the total parameters. Conversely, ResNet-50 + LoRA improves test accuracy by 14.11 percentage points over its full fine-tuning baseline, although its performance remains substantially below that of Swin-T + LoRA, indicating that LoRA effectiveness varies across architectural paradigms. These findings demonstrate that parameter-efficient fine-tuning, particularly when combined with Transformer attention layers, offers a promising approach for developing accurate and computationally efficient deepfake detection systems under resource constraints.]]>
