Alert button

Improved Detection of Face Presentation Attacks Using Image Decomposition

Mar 22, 2021
Shlok Kumar Mishra, Kuntal Sengupta, Max Horowitz-Gelb, Wen-Sheng Chu, Sofien Bouaziz, David Jacobs

Figure 1 for Improved Detection of Face Presentation Attacks Using Image Decomposition
Figure 2 for Improved Detection of Face Presentation Attacks Using Image Decomposition
Figure 3 for Improved Detection of Face Presentation Attacks Using Image Decomposition
Figure 4 for Improved Detection of Face Presentation Attacks Using Image Decomposition

Share this with someone who'll enjoy it:

Presentation attack detection (PAD) is a critical component in secure face authentication. We present a PAD algorithm to distinguish face spoofs generated by a photograph of a subject from live images. Our method uses an image decomposition network to extract albedo and normal. The domain gap between the real and spoof face images leads to easily identifiable differences, especially between the recovered albedo maps. We enhance this domain gap by retraining existing methods using supervised contrastive loss. We present empirical and theoretical analysis that demonstrates that the contrast and lighting effects can play a significant role in PAD; these show up particularly in the recovered albedo. Finally, we demonstrate that by combining all of these methods we achieve state-of-the-art results on datasets such as CelebA-Spoof, OULU and CASIA-SURF.

View paper onarxiv icon

Share this with someone who'll enjoy it: