Advancements in artificial intelligence and deep learning-based generative models have enabled the creation of near-perfect fake images called deepfakes. Deepfake images pose a certain risk in education, e-commerce, and healthcare, leading to the spread of misinformation, low-quality products, and insurance claim fraud, respectively. There are several deepfake detection models in the literature that can detect deepfake images. However, most of them require many parameters and fail to capture optimal performance in low-quality images. In this paper, we explore the lightweight architecture called MobileViT, which requires fewer parameters. We compare the model’s performance with state-of-the-art and evaluate the effectiveness of low-quality images using the Deepfake dataset. Experimental results conclude that the model successfully achieves comparable performance with fewer parameters and even generalizes well on compressed or low-quality images.