Catch up on the latest AI articles

Improving The Generalization Performance Of Face Spoofing Detection With A Simple Mixed Network, MixNet!

Improving The Generalization Performance Of Face Spoofing Detection With A Simple Mixed Network, MixNet!

Face Recognition

3 main points 
✔️ Proposed a MixNet that combines sub-networks dedicated to the detection of specific spoofing methods
✔️ Achieves higher generalization performance than models built on a single network because multiple networks complement each other
✔️ A simple and scalable framework for combining multiple networks

MixNet for Generalized Face Presentation Attack Detection
written by 
Nilay Sanghvi, Sushant Kumar Singh, Akshay Agarwal, Mayank Vatsa, Richa Singh
(Submitted on 25 Oct 2020)
Comments: Accepted at ICPR2020
Subjects: Computer Vision and Pattern Recognition (cs.CV); Cryptography and Security (cs.CR)
 
  

outline

Despite the increasing adoption of facial recognition, there are still security concerns. One of these concerns is spoofing.

There are three main types of Spoofing. The first one is Print Attack. This is a method of printing someone else's face image and using it for facial recognition. The second is Replay Attack. This is a method of displaying another person's image or video on a device such as a smartphone and using it for face recognition. The third is Mask Attack. This is a method of creating a 3D mask of another person and using it for face recognition.

When using facial recognition for airport entry/exit or office entry/exit, there are surveillance cameras and eyes around, so the possibility of these Spoofing occurring is low, and you may not be able to imagine it, but However, in applications such as eKYC, where a person's identity can be verified anywhere with a smartphone, the possibility of spoofing is high because there are no eyes around. In order to enjoy the convenience of biometric authentication and to use it safely and securely, it is necessary to support this spoofing.

Spoofing has been studied in the past and high performance has been reported. However, robustness has been an issue. In other words, the challenge is how to build a model that can cope with a variety of spoofing methods.

However. most of the algorithms to date have treated it as a binary classification task, and while they are able to determine whether or not a person is being spoofed, they are not able to learn the essential patterns of spoofing. In addition. Spoofing has different characteristics for different types of spoofing, so it is difficult for a single network to learn all these characteristics and detect them with high accuracy.

For example, I think it's obvious that Print Attack and Mask Attack have different characteristics: Print Attack is made of paper, so it has a hard surface with no unevenness and a glossy finish compared to a real face. Mask Attack, on the other hand, is more like a real face, with unevenness and a surface that is smoother and more skin-like than paper. Therefore, in this paper, we build a more robust Anti-spoofing model using a simple method that combines sub-networks of binary classifications dedicated to specific spoofing detection. In particular, we examine typical spoofing methods such as Print Attack, Replay Attack, and Mask Attack.

In this paper, we have applied the same network to all the sub-networks, but it is possible to apply the most advanced networks of Print, Replay and Mask respectively, which is a highly scalable model. We also believe that this framework can be extended not only to the field of face recognition but also to other biometrics such as iris and fingerprint.

To read more,

Please register with AI-SCHOLAR.

Sign up for free in 1 minute

OR
  • メルマガ登録(ver
  • ライター
  • エンジニア_大募集!!
Takumu avatar
I have worked as a Project Manager/Product Manager and Researcher at internet advertising companies (DSP, DMP, etc.) and machine learning startups. Currently, I am a Product Manager for new business at an IT company. I also plan services utilizing data and machine learning, and conduct seminars related to machine learning and mathematics.

If you have any suggestions for improvement of the content of the article,
please contact the AI-SCHOLAR editorial team through the contact form.

Contact Us