A Study of Bias in Gender and Racial Classification From Face Images Using FaceNet

This study investigates the challenges of bias and performance in gender and racial classification using facial recognition techniques. Existing models like FaceNet, while effective for facial verification, lack optimization for multitask classification, particularly in addressing bias across demographic groups. To explore this, we customized the FaceNet model by freezing its convolutional layers for efficient feature extraction and adding a fully connected layer for gender and racial classification. Using the UTKFace and FairFace datasets, we evaluate not only classification accuracy but also the model's fairness across different demographic categories. CFNet achieves 92% accuracy for gender 79% for race on UTKFace and 92% for gender and 65% for race on FairFace. These results reveal strengths in gender classification but highlight challenges in handling underrepresented racial groups, emphasizing the role of balanced datasets and model adjustments. This work provides a foundation for mitigating bias and enhancing multitask classification performance on resource-constrained devices.