Deepfake video detection using CNN and RNN with OPTICAL FLOW features
Elkapally Mani Sathwik Reddy, Aitipamula Pavan Kumar, Polisetty Swetha · 2024
Recent developments in machine learning have produced new technologies that make it simple to produce "deepfake" videos—videos with convincing face swaps and minimal evidence of editing. It's easy to imagine scenarios in which these realistic fake videos are exploited to cause violent protests, blackmail someone, or fabricate terrorist incidents. Digital content that has been synthesized is used to create extremely realistic-looking fake videos that fool viewers. Generative Adversarial Networks (GAN), a type of deep generative algorithm, are frequently used to do such tasks. By using this technique, realistic contents are synthesized that are highly challenging for conventional detection techniques to identify. Most of the time, discriminators based on convolutional neural networks (CNNs) are used to identify such modified media. Since the technique primarily concentrates on the spatial characteristics of each frame of video and is unable to learn time-related data obtained from inter-frame interactions, we utilized an optical flow-based feature extraction approach to extract time-related features, which are then used for classification. The foundation of this model is the integration of RNN and CNN with optical flow feature architecture.