Mobilefacenet not working good with South East Asian face

Hi, i'm currently having an attendance project and working with mobilefacenet architecture. I used faces_emore dataset. Although the accuracy for eval set is pretty high (98~99%) but still my model is having trouble recognize people and often make wrong decision. After testing with multiple cases, I have draw a conclusion that It's not good enough for Asian people. What shoud I do to improve it? Is there anyway? Thanks :D