The rapid evolution of deepfake technology, driven primarily by Generative Adversarial Networks (GANs), has transformed media, learning, and communication. However, this advancement also raises significant concerns regarding misinformation and security. Traditional detection systems, such as LSTM networks for temporal inconsistency analysis and ResNext models with advanced feature extraction, have made strides in identifying deepfakes. Nonetheless, the production of increasingly high-quality fakes, insufficient training data, and the need for real-time detection remain formidable challenges. This paper introduces a novel approach to deepfake detection, leveraging deepfake images themselves to train GANs. By utilizing deepfake-generated images as training data, the GAN is exposed to a broader range of synthetic variations, enhancing its ability to detect even the most sophisticated deepfakes. Our paper demonstrates the limitations and scope for improvement of this method.