Detection of Deepfake Videos Using a Deep Learning Approach with Res-Next and LSTM

With the increasing computational power, deep learning algorithms are significantly more complex. Hence, it became possible to generate highly realistic videos, created by AI, termed deepfakes. The videos are especially face-swapped deepfakes and are very risky because they may be employed for some malevolent intentions, for instance, in political processes, sharing fake news and other things. Therefore, the threat is significant, and so is the need for proper detection methodologies. This project concerns the differentiation between a real and a deepfake using a novel approach in deep learning. In our solution, AI is seen to counterbalance the malicious use of AI technologies, especially in creating deepfakes. Both replacement and recreation deepfakes can be automatically detected by the system. Its frame-level feature extraction was performed through a Res-Next Convolutional neural network that followed the architecture of Long Short Term Memory based Recurrent Neural Network. That enables it to organize videos into real or deepfakes that capture both spatial features of the frames and temporal aspects. So as to make it strong in practice, then used the model on a big and balanced dataset. Our project combined data coming from Face-Forensics++, the Deepfake Detection Challenge, and Celeb-DF. Results of this idea demonstrating that the suggested model has competitive performance and acts like an efficient and reliable solution for deep fake detection in real-time with simple and effective architecture.