This thesis investigates the role of data quality in achieving robustness fairness in machine learning models. Instead of evaluating traditional accuracy-based fairness, this research focuses on how a model's resilience to natural noise varies across different subgroups and aims to quantify how individual data points contribute to these robustness disparities. The study will develop methods to assess sample importance relative to subgroup stability, exploring the tension between data that maximizes overall accuracy versus data that harmonizes robust performance across minorities. Through empirical validation across diverse architectures, this work will provide data curation strategies to ensure AI models are not just universally robust, but equitably dependable.