1. Paper information
论文名称：Exposing GAN-generated faces using inconsistent corneal specular highlights
Article link: https://arxiv.org/pdf/2009.11924.pdf
Author team:
Conference: 2021-ICASSP
2. Innovation
With the help of physiological signals, a scheme is proposed to study whether the highlight patterns on the corneas of the two eyes are consistent to judge whether it is a Deepfake picture. The left picture in the figure below is a real picture. The highlight areas obtained in the left and right corneas are very similar, while the right picture is a Deepfake picture, and the highlight patterns are obviously different.
3. Method
-
The specular patterns on the corneas of the two eyeballs are correlated, determined by: interpupillary distance, diameter of the cornea, position of the two eyeballs relative to the lens (determined by head orientation), distance from the light source to the eye.
-
The analysis environment mainly mentioned in this article is the same as that required to take a normal ID photo. Steps: 1) First lock the face position, then use landmark to extract the eye position 2) Then extract the cornea position 3) Then extract the highlight pattern, use Adaptive threshold filtering is enough 4) Then the highlight patterns extracted from the two eyes are corrected, following the formula:
-
Among them, RL refers to the highlight area of the left eye, and RR refers to the highlight area of the right eye. The numerator part of the formula is the overlapping part of the highlight, and the denominator is the union of the highlights. The larger the value, the greater the similarity. The more likely it is the real face.
Four. Summary
1) Advantages: The method is interpretable; the method relies on physiological characteristics, which is more robust and generalizable than relying on the semantic features of images. 2) Disadvantages: There are high requirements for the orientation of the face, lighting, and clarity; and the inconsistency of these highlights can be eliminated through post-processing.