Balancing race, gender, and age accuracy in face recognition using DB-BETA-VAE

This paper addresses the multi-dimensional bias issues (race, gender, age) in face recognition systems by proposing a debiasing method based on an improved DB-BETA-VAE model. By integrating latent space resampling with a multi-task joint training mechanism, the model achieves simultaneous optimization of accuracy and fairness in race, gender, and age recognition on the UTKFace dataset. Experimental results demonstrate that DB-BETA-VAE attains an average accuracy of 0.83 in racial classification (DIR=0.85, EOD=0.15), significantly outperforming traditional CNNs (average accuracy: 0.78, DIR $=0.526$, EOD $=-0.37$) and comparative models RL-RBN (0.68) and FairMixRep (0.65). It shows particularly strong performance for minority groups (e.g., 0.60 accuracy for “Other races”) and older demographics (0.45 for ages 50-59).Additionally, the model effectively mitigates gender bias, reducing classification disparities to an average of 0.03 (e.g., $0.91 / 0.87$ accuracy for White males/females), while exhibiting optimal stability in age prediction (${\mathrm {S D}}=0.205$). The study confirms that DB-BETA-VAE’s combination of feature disentanglement and dynamic sampling strategies significantly reduces multi-dimensional biases while improving overall performance, offering a practical solution for AI fairness research.