Meta-learning for efficient few-shot classification in facial liveness detection | |
| Author | Htoo Lwin |
| Call Number | AIT RSPR no.CS-21-05 |
| Subject(s) | Computational intelligence Biometric identification |
| Note | A research study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Computer Science |
| Publisher | Asian Institute of Technology |
| Abstract | Facial liveness detection - also known as face anti-spoofing (FAS) - is an increasingly crucial component of robust real-world facial recognition systems. Recently, deep learning models have achieved remarkable success in identifying “spoofs” across various benchmarks, but the diverse array of spoof types that can be devised against such systems is impossible to an ticipate and train against. Even if training samples for new spoof types were available, their number would be few, and collecting a sufficiently large number for ordinary supervised learning techniques would be expensive and time-consuming. Meta-learning is a possible alternative approach to this few-shot classification problem. This study evaluates the effec tiveness of Model-Agnostic Meta-Learning (MAML) in two scenarios: a “known-spoof” scenario in which all spoof types are known and an “unknown” scenario in which the model must be adapted to a new previously unseen spoof type after obtaining one or a few ex amples. The results show that meta-learning (MAML) outperforms traditional models on unseen FAS data, while traditional models outperform meta-learned models on seen FAS data. Therefore, on the onset of an unseen spoof type, it is recommended to use MAML as a stopgap model while sufficient data is being collected to re-train the traditional model. Another recommendation is to use this stopgap model as part of an ensemble committee in tandem with traditional models to combine the best of both worlds. |
| Year | 2021 |
| Type | Research Study Project Report (RSPR) |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Dailey, Mathew N. |
| Examination Committee(s) | Mongkol Ekpanyapong |
| Scholarship Donor(s) | AIT Fellowship |
| Degree | Research Studies Project Report (M. Eng.) - Asian Institute of Technology, 2021 |