FairSAM: Fair Classification on Corrupted Image Data Through Sharpness-Aware Minimization

FairSAM is a new framework that integrates fairness-oriented strategies into Sharpness-Aware Minimization (SAM) to address the performance degradation of image classification models across demographic subgroups when exposed to corrupted data. It introduces a metric for assessing performance disparities under data corruption and demonstrates through experiments on various real-world datasets that FairSAM effectively balances robustness and fairness. This development is significant for practitioners as it provides a structured approach to mitigate algorithmic bias while maintaining model performance in adverse conditions.