Improved AI-Based Facial Recognition System to Mitigate Demographic Biasness

Facial recognition systems have achieved remarkable success in recent years; however, they continue to suffer from demographic bias, often leading to unequal performance across different population groups. This study presents a novel bias mitigation framework that simultaneously improves both accuracy and fairness, challenging the widely held assumption of a trade-off between these two objectives. The proposed approach integrates uncertainty-based sample selection, human-in-the-loop expert labelling, and continual learning with prior preservation to address bias in a targeted and adaptive manner.

The framework focuses on identifying and correcting uncertain predictions, which are frequently associated with underrepresented groups, thereby enhancing performance without degrading accuracy on well-represented populations. Experimental results demonstrate the effectiveness of the proposed method. On the FairFace dataset, accuracy improved from 92.58% to 96.47%, while the Degree of Bias (DoB) was reduced from 3.86 to 1.16, representing a reduction of approximately 70%.

Notably, significant performance gains were observed for underrepresented groups, such as an improvement in accuracy for black females from 80.2% to 90.86%, while maintaining strong performance for already well-represented groups. These findings confirm that fairness and accuracy can be improved simultaneously through targeted, data-driven interventions. The proposed framework offers a practical and scalable solution for developing more equitable and reliable facial recognition systems suitable for real-world deployment