A Hybrid Deep Learning Model for Robust Deepfake Detection

The emergence of deepfake videos has become a major concern due to their capability to mislead and deceive viewers. This paper presents a novel deep learning framework that combines the strengths Generative Adversarial Networks (GAN) and Capsule Networks (CapsNets). The goal of the project is to provide an improved and resilient framework against attacks, and to achieve superior accuracy. The study examines the existing techniques for detecting deepfake videos. It also explores the limitations of single models, such as CNNs and LSTMs. The research finds inspiration in the GANs’ ability to produce synthetic deepfake samples, as well as in CapsNets’ ability to extract hierarchical relationships. A hybrid framework is then developed to combine these features. The proposed hybrid framework enhances the training dataset by incorporating GANs’ ability to produce synthetic images, and the CapsNet's expertise in spatial relationships. As a classifier, it is trained using the latter's results, and a decision-level fusion procedure is carried out to increase its robustness. The evaluated datasets consisted of real and synthetic deepfake videos. The results of the experiments revealed that the hybrid model was able to achieve a 96.16% accuracy rate, which is higher than the accuracy of CNN and LSTM models. The impressive performance of the hybrid model highlights how it can distinguish genuine and manipulated content. The proposed model was also evaluated against different evaluation parameters, such as recall, precision, AUC-PR, and F1 score. Consistent with its robust and accurate capabilities, the hybrid model excelled in all of these tests. The research presents a hybrid framework that combines the strengths of both GANs and the CapsNets. It was able to achieve an accuracy rate of 96.16%, which is significantly higher than the accuracy of CNN or LSTM models. The findings of this study encourage the development of more effective and reliable deepfake detection systems.