Bias Mitigation in Anti-Spoofing through Knowledge Distillation

In the same way that face recognition algorithms suffer from bias due to the way they have been trained, anti-spoofing algorithms may undergo such limitations if the bias is not mitigated during the training phase. To make algorithms as robust as possible towards human diversity Unissey uses knowledge distillation. It is a technique enabling a student model to learn the generalization power of a cumbersome teacher. Here, Unissey presents a distillation process where multiple experts trained to be excellent on their respective domain teach a student model to mimic their individual performances and eventually become excellent on every domain. Unissey shows how this technique can be used to mitigate bias related to gender and ethnicity.

Curriculum vitæ

Idriss Mghabbaris currently a computer vision engineer at Unissey where he is working on presentation attack detection. Previously Idriss interned at BNP Paribas, the largest French banking group, where he worked on machine translation using neural networks. Idriss graduated from Centrale Paris and Ecole Normale Supérieure of Cachan.