The key component of contemporary biometric security systems is the accurate identification and validation of individuals using facial recognition technology. Advanced security systems can now be personalized through improved integration of gender recognition in their development. The face recognition models VGG16, VGG19 and ResNet50 together with other identifiers face issues to adapt to variations in skin tone specifically affecting users with darker skin complexion. The ability of these models strongly depends on skin melanin levels, which results in higher cases of false identification or bias towards users with these characteristics. Progress has been made, but the deficiency of skin-type representation during training con- tinues to result in performance discrepancies among numerous operational systems. The common method for evaluating and training face recognition models is based on machine learning technology. These face recognition datasets must have large- scale deployments of wide skin tone diversity and multiple age ranges to improve model recognition capabilities independently of demographic features. Transfer learning solutions create better model accuracy along with fairness, since they help large dataset- trained models adapt to small heterogeneous datasets. This method proves beneficial in the identification of human facial features for biometric validation because it showcases enhanced performance and shorter training periods, as well as universal real-world scenario applicability. The proposed work is suitable for practical real-world applications in Airport security, Border Control along with Border Surveillance and Law enforcement and finance, Healthcare, Mobile devices and security systems. In order to ensure more reliable and inclusive face recognition solutions, biometric systems can use transfer learning to overcome the challenges posed by diverse facial characteristics and skin tones. The results obtained from the experimental evaluation carried out on the UTKFace dataset show that the best performance is achieved by the ResNet50V2 model, which has the highest accuracy of 84% in the validation set compared to the accuracy achieved by the other architectures.
K et al. (Wed,) studied this question.