Facial Password Data Augmentation

Abstract

We present a series of data augmentation methods that utilize noise in a positive test case as multiple negative test cases. These data augmentation methods are utilized to increase the accuracy of a facial authentication system that uses facial motion concurrently with conventional facial identification to verify a person’s identity. We propose using non moving frames from a video of a facial motion, as negative cases to increase the significance of the negative cases used during training. We will also use single frames repeated from the motion to supply more negative samples of non-moving faces. These methods are useful in training the network to distinguish a facial motion from a non moving face. The data augmentation will also be used during evaluation of the network to assign each facial motion password a strength value based on how they compare to the augmented data.