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

Real-world AI systems face a dual fairness problem: When image classification models encounter corrupted or degraded data (like blurry or noisy images), they don't just perform worse overall—they perform disproportionately worse for certain demographic groups, amplifying algorithmic bias in deployment scenarios.

Existing fairness and robustness solutions work against each other: The paper identifies a fundamental tension where methods that improve overall robustness don't address fairness disparities, and fairness-focused approaches fail to maintain equitable performance when data quality degrades, leaving a critical gap in responsible AI.

FairSAM provides a practical framework for balanced performance: By combining fairness strategies with Sharpness-Aware Minimization (a robustness technique), the proposed approach successfully maintains more equitable accuracy across demographic groups even when images are corrupted, demonstrating that robustness and fairness can be jointly optimized.

The problem is measurable and addressable: The researchers introduced a new metric to track performance degradation across demographic subgroups under data corruption, enabling practitioners to identify and monitor this specific fairness risk in their own systems.

This matters for real-world deployment: Testing on multiple real-world datasets confirms the framework works in practice, suggesting organizations can implement fairer and more robust image classification systems without accepting the current trade-off between overall performance and equitable treatment across demographic groups.