Design and Analytical Performance of a Hybrid Model Using Deep Learning Techniques for the Detection of Deepfake Videos

Recently, it has been possible to develop utilities that can create video face-swaps that are indistinguishable from the real. Recent advancements in deep learning techniques make it possible to synthesize ever-realistic videos. Popularly referred to as deepfakes, this technology has indeed brought a lot of ease and more realistic fakes than the normal ones. In as much as deepfakes are easy to make, their detection is challenging. The study presents a hybrid model that integrates GRU in EfficientNet-B1, InceptionV3, Xception, and DenseNet121. These models were evaluated as pure networks without the GRU layers and as networks with the GRU layers positioned to process temporal information. The GRU, which is a recurrent neural network variant, serves the purpose of adding an extra edge in performance in detecting deepfakes by looking at the temporal aspect of the data. Over the course of the research, there has been an emphasis on standard models and their variation with GRU models in order to understand the role of time-based data. This study enhances constant efforts against deepfakes and will help researchers & practitioners who develop deepfake detection systems and protect the integrity of the digital media.