Date of Award
2025
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
Committee Chair
Tathagata Mukherjee
Committee Member
Letha Etzkorn
Committee Member
Chaity Banerjee Muhkerjee
Research Advisor
Tathagata Mukherjee
Subject(s)
Human face recognition (Computer science), Artificial intelligence--Moral and ethical aspects, Machine learning
Abstract
Facial recognition technology is widely used in security and identity verification but exhibits biases that disproportionately impact certain demographic groups. This thesis examines racial and skin tone biases in commonly used facial recognition datasets, assessing their influence on classification models. Using manual classification via the Fitzpatrick scale and numerical skin tone representations from LAB color values, we identified significant dataset imbalances, with lighter skin tones (Fitzpatrick Types 1 and 2) overrepresented and darker tones (Types 4-6) underrepresented. Model evaluation using Random Forest, SVM, and XGBoost showed an overall accuracy of 72\%, but classification for Asian individuals was notably weaker (0.45 precision, 0.36 recall), suggesting dataset bias affects performance. Comparisons with US Census 2020 data further revealed discrepancies in racial representation, underscoring the need for better dataset alignment. To mitigate these issues, this study highlights the importance of manual auditing and human-in-the-loop methodologies in ensuring fairness. Future research should focus on curating balanced datasets, employing fairness-aware algorithms, and analyzing bias propagation within facial recognition models to advance equitable AI development.
Recommended Citation
Williams, MilYonta, "Evaluating bias in facial recognition datasets : a study on representation and classification fairness" (2025). Theses. 751.
https://louis.uah.edu/uah-theses/751
