Face spoofing detection
There has been a growing adoption of face recognition systems being used for biometric authentication. Therefore, it is crucial to ensure that facial recognition systems are safe from malicious impersonation attempts to gain access. Which could to theft of confidential information or valuables store...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2022
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/157823 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | There has been a growing adoption of face recognition systems being used for biometric authentication. Therefore, it is crucial to ensure that facial recognition systems are safe from malicious impersonation attempts to gain access. Which could to theft of confidential information or valuables stored in applications these systems are typing to protect
Thus, this project proposes and develop a presentation attack detection system based on a lightweight machine learning model for Near Infrared (NIR) camera facial recognition systems such that it could work well on devices with limited computational power such as mobile phones.
The proposed model will classify faces detected from a face detection model as a live or spoof face. As there are limited face spoofing detection datasets with NIR image, this project opted to use RGB datasets (CelebA-Spoof and LCC_FASD) that are easier to find, and data augment the images to look more like NIR images. The proposed model will focus on being lightweight while trying to maintain state of the art performance in detecting impersonation attempts.
Lastly while the proposed model demonstrates strong performance during the testing, with achieving an accuracy of 95.6% and 84.5% when evaluated with the validation and test dataset respectively with the model size being as low as 2MB. This report also discusses the results from the live demonstration to test the model’s strengths across different real-life scenarios. |
---|