Detecting AI-Generated Deep Fakes Using ResNext CNN and LSTM-Based RNN: A Robust Approach for Real-Time Video Manipulation Detection

The increasing computational power has greatly empowered deep learning algorithms, making the creation of highly realistic and virtually undetectable synthetic videos, commonly known as deep fakes, remarkably simple. These deep fakes, often featuring convincing face swaps, can be employed in situations such as generating political turmoil, fabricating terrorism events, distributing revenge porn, and engaging in blackmail. In this study, we introduce an innovative deep learning-based technique designed to effectively differentiate between AI-generated fake videos and genuine ones. Our approach excels at automatically identifying manipulated videos, including those involving content replacement and reenactment deep fakes. Our endeavor involves leveraging Artificial Intelligence (AI) to combat its own creations. The central component of our system utilizes a ResNext Convolutional Neural Network (CNN) to extract features from individual frames of videos. These features are then used to train a Long Short-Term Memory (LSTM)-based Recurrent Neural Network (RNN). This RNN helps classify videos by determining if they have been manipulated, such as being deepfakes, or if they are genuine recordings. To ensure our model performs well with real-time data and reflects real-world situations, we evaluate it on a large and balanced dataset created by combining various existing datasets like FaceForensic++, Deepfake Detection Challenge, and Celeb-DF. Furthermore, we showcase that our system achieves competitive results using a straightforward yet effective approach.