Deepfake Video Detection: Analysis for Deep Learning models using Transfer Learning

The unchecked spread of deepfake videos jeopardizes public trust, national security, and media credibility. The capacity to recognize falsified videos is critical for protecting privacy and combatting misinformation. The article seeks to illustrate the usage of Deep Learning (DL) approaches for recognizing deepfake videos, including Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and an RNN-LSTM combination with Transfer Learning. The research used the LSTM model to examine the efficacy of temporal correlations between video frames in the Deepfake Detection Challenge (DFDC) dataset of deep fake videos, and it claimed complete success in both the training and testing stages. In addition, the study used frame level masks and characteristics that were created throughout the research's pre-processing steps. The findings show that the model had a train loss of 0.021 and a test loss of 0.813, suggesting that it was extremely resilient. Other crucial indicator measures, such as Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM), aid in evaluating the model's performance and validate its high accuracy in recognizing deepfake material. This study is very important since it greatly contributes to the body of knowledge in the early phases of deepfake identification and aids in the mitigation of the risks connected with manipulated media.