Deep Fake Video Classification with Sequential Input Frames Using Hybrid Deep Learning Model and Bayesian Optimization

With the advancement of Artificial Intelligence, creating synthetic content has become effortless. Manipulated content, such as synthesized images and videos known as Deep-fakes, is now prevalent on the internet and social media platforms. These deepfakes are generated using deep learning techniques, such as Generative Adversarial Networks (GANs) or Autoen-coders (AEs). They have the potential to spread false information and are particularly concerning in cases involving identity theft, especially when involving well-known celebrities. Distinguishing these fake videos with the human eye is challenging, as they closely resemble real ones. This research utilizes popular deep learning (DL) models, such as CNN (VGG19) combined with stacked LSTM, for the classification of real and fake videos. Sequences of input frames are stacked and fed into the hybrid model, which is developed using the TensorFlow framework for this study. Our approach applies video-level classification rather than frame-level classification, and it utilizes Bayesian Optimization (BO) for hyperparameter tuning, distinguishing it from other recent methods. The VGG19-LSTM model is trained and tested against FaceForensics dataset(FF++) to determine the best parameters and achieved the highest test accuracy of 95.6% for deepfakes dataset.