A Machine Learning Based Approach for Deepfake Detection in Social Media Through Key Video Frame Extraction

摘要: In the last few years, with advent of deepfake videos, image forgery has become a serious threat. video, person’s face, emotion or speech are replaced by someone else’s different speech, using deep learning technology. These videos often so sophisticated that traces manipulation difficult to detect. They can have heavy social, political and emotional impact on individuals, as well society. Social media most common targets they vulnerable platforms, susceptible blackmailing defaming person. There some existing works for detecting but very attempts been made in social media. The first step preempt such misleading from is detect them. Our paper presents novel neural network-based method fake videos. We applied key video frame extraction technique reduce computation A model, consisting convolutional network (CNN) classifier network, proposed along algorithm. Xception net chosen over two other structures—InceptionV3 Resnet50—for pairing our classifier. model visual artifact-based detection technique. feature vectors CNN module used input subsequent classifying video. FaceForensics++ Deepfake Detection Challenge datasets reach best model. detects highly compressed high accuracy lowered computational requirements. achieved 98.5% dataset 92.33% combined Challenge. Any autoencoder generated be detected almost all if possess more than one frame. reported here when number frames one. simplicity will help people check authenticity work focused, not limited, addressing economical issues due this paper, we achieve without training an enormous amount data. reduces computations significantly, compared works.