Deepfake videos can imitate the visual appearance of any person, including celebrities and famous personalities. Today we observe many untestified videos of famous persons on internet. We aim to solve the problem of deepfake videos by using the domain knowledge about the data distribution of the famous person as well as the procedures that produce the deepfake videos.

Fig 1. Distribution of statistic for real and fake images.

Fig 2. Example of fake images generated by deep neural network.

Fig 3. Separation of real and fake images

Celebdf id: 0
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.97
Celebdf id: 1
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.99
Celebdf id: 2
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.99
Celebdf id: 3
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.86
Celebdf id: 4
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.71
Celebdf id: 5
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.81
Celebdf id: 6
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.97
Celebdf id: 7
Test_v1_[0, 1], Test_v2_[0, 1] AUC= 0.92