Liveness Detection

The client required an enhancement in the metrics of their biometric access system for banking programs; the neural network was missing 18% of attacks. According to specifications, the anti-spoofing dataset necessitated a diverse array of mobile devices and accessories (masks, wigs, glasses, scarves, etc.). The client approached us due to their internal team’s unsuccessful experience in data collection, coupled with time constraints regarding the dataset

Since the biometric system only required refinement for complex and atypical appearance scenarios, we artificially generated 10,000 video attacks. We assembled a team of 20 individuals who created diverse filming conditions, using paper masks and various props. Each assessor simulated attacks involving 500 individuals

• The client successfully trained the neural network, reducing the number of missed attacks from 18% to 1%
• A pre-assembled “bank of props” tailored for such tasks enabled us to complete the project 3 days ahead of schedule