Deepfake Facial Recognition for Video Clips

The accessibility of deep learning tools has fueled the rise of convincing Deepfake videos. This research focuses on developing and evaluating an advanced Deepfake Facial Recognition system, integrating Long Short-Term Memory (LSTM) and ResNeXt CNNs. Utilizing CNNs for frame-level facial feature extraction and LSTM for temporal dependencies, the system excels in discerning nuanced manipulations. Extensively tested on diverse Deepfake content, our findings demonstrate the system's considerable accuracy in distinguishing between authentic and manipulated facial features. The integrated LSTM enhances sensitivity to subtle manipulations often overlooked by frame-based methods, while the ResNeXt CNN architecture improves overall precision. Rigorously evaluated on a diverse set of Deepfake-containing video clips, the system proves highly effective. The research also addresses challenges in Deepfake facial recognition, exploring potential countermeasures and contributing valuable insights to the development of robust systems mitigating risks associated with manipulated video content. The FaceForensic++ dataset model stands out with its remarkable precision, achieving an impressive accuracy rate of 97.76%. Meanwhile, the Celeb-DF dataset model has also performed admirably, boasting a competitive accuracy score of 93.97%.