Preserving the Veracity of Digital Media using Neural Network Based Detection of Deepfake Videos
Pranav Bire, Om Ambalkar, Ritesh Bagade, Pradnya Samit Mehta, Anuradha Yenkikar · 2024
Free deep learning applications have made it easier to create similar to humans synthesized videos in the past couple of years, a trend known as "deep fakes." Over many years, it has been possible to manipulate digital videos through the skilful use of visual effects. However, the simplicity with which fake content can be created and its realism have both increased dramatically due to recent advancements in deep learning. It is easy to imagine scenarios in which people are blackmailed, political unrest is caused, or events involving terrorism are fabricated using these realistic face-swapping deepfakes. This project proposal describes an exciting deep learning-based method that effectively recognizes real videos from ones produced by artificial intelligence. The use of artificial intelligence (AI) to combat AI is mentioned in the proposed model. The proposed method extracts frame-level features utilizing a Res-Next Convolution neural network. These attributes are then used to train an LSTM-based Recurrent Neural Network (RNN) to identify videos based on whether they have been altered or not, i.e., whether they are deepfake or real. The Deepfake Detection Challenge and Celeb-DF are two examples of the many available datasets that are combined to create a big, balanced, and mixed dataset that the model's performance on real-time data and apply it to real-world scenarios.