Reproducibility report for "On Disentangling Spoof Trace forGeneric Face Anti-Spoofing"

Labiba Kanij Rupty, Md. Sourave Hossain, Nabeel Mohammed

Abstract:

Scope of Reproducibility
From previous research(1), it has been demonstrated that the face anti-spoofing problem can be treated as a denoising problem where spoof images can be disentangled into two parts – the live counterpart and the noise. Based on that, this paper(2) suggests that the disentanglement will be more appropriate if done in four parts - s, b, C, T where - s, b represent color range bias and balance bias, C denotes the smooth content patterns and, T is the high-level texture6patterns. They claim that by leveraging these features, one can find the live counterpart of any image as well as the spoof counterpart of the live inputs by warping. We identify the main scope of reproducibility is to verify whether it is8possible to generate the live counterpart from any image by subtracting the spoof trace.

Methodology
We approach this challenge in the following manner. Firstly, we have reproduced the paper in PyTorch while taking help from the paper, the authors, and the official implementation. This process took us around a month. Secondly, we checked the soundness of our PyTorch implementation by taking the same input in both our and their implementation. Later, we train on MSU SiW Protocol-1 and OULU NPU Protocol-1 to verify whether we have succeeded to reproduce their results in PyTorch. Each epoch on a single V100 took around 12 hours. Finally, we went on to propose several improvements over a few identified limitations of the original paper.

Results
While verifying the reproducibility, we outperformed the result of OULU NPU Protocol-1 as we got an ACER of 1.195% compared to their 1.9%. For MSU-SiW Protocol-1, we achieved an ACER of 0.53% while they found ACER of 0%.
Later, we verified whether it is possible to get a perfect live counterpart from any input image. We concluded that it is not always possible to get perfectly warped spoof images in many cases. Later, we proposed a few techniques to improve the generation of the live counterpart based on our observation.

What was easy
Although the method was really complicated, as the official implementation was easy to follow, we were at ease while re-implementing the architecture of the paper in PyTorch.

What was difficult
The hyperparameters for producing the metrics were not given and, as mentioned in the paper, the authors selected them by brute-forcing. We found it really time-consuming to find the best result as the method is prone to hyperparameters.

Communication with original authors
After we read the paper and started implementing the code, we had a few confusions for which we mailed the author and got help from them. Later, we suggested an inconsistency between the paper and the implementation.

Paper Url: https://openreview.net/forum?id=LQapvMnUyvE