Towards An Ethnic-Bias Free Approach For Facial Recognition

Abstract

In this study, we investigate the issue of ethnicity bias on the existing facial recognition approaches and provide a novel way to consider the ethnicity at the model training stage. This study aims to explore the ethnicity bias issue on the current state of the art facial recognition algorithm Facenet and to provide a novel image selection method to minimize the effect from the ethnicity bias without making changes to the neural network architecture. Building upon the works of Facenet, modifications were done to the image selection process which triggers the training stage of the model. Since the input data to the model for training is fed using batch by batch, the batch creation operation is modified so that each batch represents the total ethnicity distribution of the training dataset. Then the trained model is evaluated thoroughly using classification tasks, clustering attempts by comparing with the base model. From the classification tasks, the focus directed towards the metrics such as accuracy, false acceptance rate, area under the curve (AUC). In the process to create batches that have the same ethnic distribution, we have presented an ethnicity classifier that can classify a given image of a person to its relevant ethnicity. The model trained with the proposed method has shown significant improvement when considering the AUC metric which reduces the chance to misclassify a person with underrepresented ethnicity. Thus reducing the false acceptance rate which increases the reliability of the proposed model as well as minority representation. The clustering evaluation tasks revealed that the model trained with the proposed approach can form near-perfect clusters while the base model struggles to do so.