Balancing Racial Accuracy in Face Recognition with DB-BETA-VAE
Xuan Cheng, Meng Li, Xinan Xu · 2025
As face recognition technology keeps developing, we've started to notice some issues with racial bias in its accuracy. To tackle this, we need to address a few major problems: the imbalance in data, human errors, and flaws in the algorithms that lead to bias in racial classifications. In this paper, we explore a deep learning technique called DB-BETA-VAE, which is designed to help reduce this bias.This approach builds on the BetaVAE framework and works by tweaking a specific parameter (the hyperparameter β) to better separate the different variables in the data. It looks at how often samples show up in the latent space and focuses on the quality of the reconstructed images. This helps identify and retrain the samples that are less likely to occur and are overly bright, which eventually helps to improve the fairness of racial classifications across various datasets.We tested this method on well-known ethnic datasets like FairFace and UTKFace, and it seems to have broad applicability. The results indicate that our technique effectively reduces racial classification bias and can generally apply across different situations. This approach not only minimizes bias without needing human re-labeling but also opens up new ideas for achieving fairness in face classification.