Robust deepfake detection using Long Short-Term Memory networks for video authentication

Developments achieved in recent years have propelled techniques for generating and manipulating multimedia content to attain an exceptionally high degree of realism. According to a survey, 25 percent of the videos viewers watch are fake. The increasingly blurred distinction between authentic and synthetic media presents significant security concerns, with the potential for exploitation in various domains. These threats encompass the manipulation of public opinion during electoral processes, perpetration of fraudulent activities, dissemination of disinformation to discredit individuals or entities, and the facilitation of blackmail schemes. Detecting fakes is tricky and difficult for viewers who are watching them, with studies showing that over 70 percent struggle to identify them accurately. To counter this issue, we envision this project whose primary goal is to construct a model that is capable of distinguishing between deepfake and authentic videos. Our proposed model operates at the video level, analyzing entire videos at once to provide a comprehensive assessment. The dataset utilized for training and evaluation is sourced from repositories such as DFDC, FaceForensics++ and Celeb-DF. The dataset sourced from DFDC and Celeb-Df are converted into frames from videos, in this architecture first face recognition tool is used for detecting the faces, followed by ResNext for feature extraction and then LSTM is used to classify the videos.