Fake Video Detection using Modified XceptionNet
Imam Kusniadi, Arief Setyanto · 2021
Deepfakes are a significant threat to society from the development of deep learning technology, which is able to manipulate digital images. Manipulated digital content potentially become disinformation. This study proposes the task of detecting fake video using the XceptioNet architecture through transfer learning. The proposed algorithm train and test with publicly available datasets FaceForensics ++. The dataset is pre-processed to filter the face only using MTCNN. This research proposes an improvement of XceptionNet architecture with fine tuning and transfer learning. The purpose of this study is to determine the effect of architectural changes and the number of frames per video (FPV) towards the accuracy. Research shows the addition of 3 fully-connected layers in front of the softmax layer with the fine tuning lead to superior accuracy. The highest testing accuracy achieved on Celeb-DF dataset at 83,75%.