Advancements in Deepfake Detection: Leveraging Bi-LSTM-CNN Architecture for Robust Identification

In our modern era, the saying “visual evidence is conclusive” no longer holds true, presenting significant implications across various spheres of our lives. With the rapid advancement of technology, the creation of synthetic media, particularly deepfakes, has become remarkably facile. Some applications even facilitate the creation of deepfakes directly on handheld devices. The identification of deepfakes poses a formidable challenge as they can be imperceptible to the human eye. Nevertheless, researchers are actively <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sup>ngaged in developing methodologies to discern deepfakes. Deepfakes constitute media generated through AI algorithms. These algorithms assimilate attributes from a reference image and overlay them onto a source image. In our pursuit of identifying video deepfakes, we employ deep learning architectures such as Bi-LSTM integrated with CNN. Our approach capitalizes on 50 epochs of training to attain a remarkable accuracy of up to 98.7%. We leverage transfer learning to construct a robust deepfake detection model. Initially, we utilize a pretrained CNN, dubbed Bi-LSTM-CNN, to extract salient features and construct feature vectors. Subsequently, the Bi-LSTM layer is trained using these feature vectors. The resulting confusion matrix substantiates the efficacy of our model, with validation and testing accuracies reaching an impressive level.