Description

The MSU USSA database was created to simulate spoof attacks on smartphones. With the release of Android 4.0, millions of smartphones now have the ability to unlock their device using the Trusted Face (face unlock) functionality such as the Google Nexus 5. Hence, the cameras used in capturing the MSU USSA database simulate the input the Trusted Face application would receive when a malicious user tries to spoof the system. Spoof attacks captured by DSLR or USB webcams do not replicate the application of user authentication on smartphones as they contain additionally sensors (DSLR Cameras) or capture blurry out of focus images (USB webcams).

Moreover, current public-domain spoof databases lack diversity in terms of background, illumination, and image quality, and thus do not replicate real application scenarios. The MSU USSA database was also created to ensure that it contains a mixture of environments, image qualities, image capture devices and subject diversity. To create such a database we selected 1,000 live subject images of celebrities from the Weakly Labeled Face Database (http://wlfdb.stevenhoi.com). Therefore, the public set of the MSU USSA database for face spoofing consist of 9,000 images (1,000 live subject and 8,000 spoof attack) of the 1,000 subjects.

Due to IP issues, the public set we released is slightly different from the public set used in our paper (the genuine and spoof images of the subjects from the Idiap and CASIA databases are not included). In order to allow other researchers to compare their results with our results, this webpage contains spoof detection performance on the new public set of the MSU USSA database. Please note there is only a slight change in the performance results.

The database was produced at Michigan State University’s Pattern Recognition and Image Processing (PRIP) Lab, in East Lansing, MI, US.

Figure 1: Sample images of (a) live faces, (b) spoof faces captured by the front facing camera, (c) spoof faces captured by rear facing camera on the Nexus 5. The 4 spoof images are captured using 4 spoof mediums in the following order MacBook, Nexus 5, Nvidia Shield Tablet, and Printed Photo.

Details

Camera configuration

To capture the live subject images, a wide variety of cameras were used as these images where downloaded from the internet.

Two types of cameras were used in collecting the spoof attacks:

  1. Front-facing camera in the Google Nexus 5 Android phone (1280 × 960).
  2. Rear-facing camera in the Google Nexus 5 Android phone (3264 × 2448).
A highly desirable property of capturing spoof attacks with smartphone devices is that it simulates input images that may be presented to smartphones that contain facial recognition systems, such as the Google Nexus 5.

Capturing Spoofing Attacks

To generate the spoof attacks we used the Front and Rear facing camera on the Google Nexus 5. Given that most people have access to a laptop, tablet or smartphone, we captured replay attacks on all three spoof mediums. The spoof attacks are captured by showing the live face image on the screen of one of the spoof mediums and using both the front and rear facing cameras of the Google Nexus 5 to capture the simulated attack. To generate printed photo attacks, we printed images of all the subjects using a HP Color Laserjet CP6015xh printer (1200 × 600dpi) on a matte 8.5 × 11-inch white paper and again used the front and rear facing camera of the Nexus 5 to simulate the attack.

Spoof Mediums:

Protocols for Spoofing Attacks

To evaluate the performance of anti-spoofing methods on the MSU USSA database please follow the protocol described in this section.

Perform a Five-Fold Subject Exclusive Cross-Validation using the subject IDs listed in the FiveFoldSubjectID file to split up the 1,000 subjects into their respective 5 folds. Each fold will contain the one live subject image its corresponding four spoof images for a subject. For each distinct fold, use the other 4 folds to train a classifier of your choosing and testing it on the remaining fold. Performance should be reported as the average Equal Error Rate (EER) and Standard deviation of your classifier(s) performance across the 5 testing folds.

Download Instructions

To download the MSU USAA face spoof database, please first print out, fill and sign the Agreement and send it to: Debayan Deb (debdebay@msu.edu). You will receive a download link upon approval of your usage of the database.

Acknowledgements

If you use this database, please cite the following publication: 


@ARTICLE{PatelTIFS16,
author={Keyurkumar Patel and Hu Han and Jain, A.K.},
journal={IEEE Trans. Information Forensic and Security},
title={{ Secure Face Unlock: Spoof Detection on Smartphones}},
year={2016},
month={June},
volume={XX},
number={XX},
pages={XXX},
}