Methods for mitigating bias of biometric systems

Biometric analysis systems, especially those employed for soft attribute classification tasks like gender or race estimation, frequently display significant biases against specific demographic groups. This inherent unfairness compromises the integrity and equity of algorithmic decision-making processes relying on these systems. While mitigating such bias is critically important for responsible AI deployment, existing techniques often encounter substantial limitations. These include poor generalizability across different datasets or conditions, a heavy reliance on labor-intensive and privacy-sensitive annotated demographic data, and an inherent conflict where efforts to maximize fairness can detrimentally affect overall classification accuracy, particularly impacting performance for well-represented groups. Addressing this complex challenge, this research provides a detailed investigation into the manifestation of these biases within soft biometric algorithms. More importantly, it puts forth innovative mitigation frameworks, exploring techniques such as deep generative models and self-supervised learning paradigms. The central aim of this dissertation is to achieve a more effective balance between equity and utility. It seeks to substantially improve fairness metrics across diverse demographic groups while crucially avoiding significant detriment to the overall accuracy and generalization performance of the biometric system, paving the way for more trustworthy and equitable applications.

Author

• Ramachandran, Sreeraj

Identifier

hdl:10057/30064

OAI identifier oai:identifier

oai:soar.wichita.edu:10057/30064