Dataset Protocols

Biometric Recognition

The biometric recognition protocol for both face and voice is grandtest0-voice-bio. In this protocol, session 2 data is used for enrollment and all other sessions are used for probing. The IDIAP and MPH-FRA are used for development and NTNU and MPH-IND are used for evaluation. Voice videos are used for both face and voice. That is why the protocol has the word voice in its name.

Vulnerability Analysis

The vulnerability analysis protocols consist of evaluating a biometric system twice: once using only bona-fide samples consisting of genuines and zero-effort-impostors and using only bona-fide genuines (no zero-effort-impostors) and presentation attacks. The protocols are called licit and spoof, respectively. The grandtest1-voice-licit protocol is used for both face and voice modalities. The licit protocol is very similar to the biometric recognition protocol except that only session 1 is used for probing. Only session 1 is used in probing because the presentation attacks are created from session 1 and this allows a comparison with minimal differences between bona-fide and presentation attacks.

grandtest1-voice-spoof and grandtest1-face-spoof are used as spoof protocols of voice and face modalities, respectively. All presentation attacks are used in spoof protocols as probes.

Presentation Attack Detection

PAD protocols are created according to the SWAN-PAD-protocols document. Bona-fide session 2 data is split into 3 sets of training, development, and evaluation. The bona-fide data from sessions 3,4,5,6 are used for evaluation as well. PA samples are randomly split into 3 sets of training, development, and evaluation. All the random splits are done 10 times to created 10 different protocols. The PAD protocols contain only one type of attacks. For convenience, PA_F and PA_V protocols are created for face and voice, respectively which contain all the attacks.