Diff-Fair: Mitigating Intersectional Bias Through Diffusion-Driven Fair Representation

ICLR 2026 Conference Submission13575 Authors

Keywords: Data Fairness, Intersectional Bias, Multiple Sensitive Attributes, Fair Representation

Abstract: Algorithmic fairness remains a critical challenge in Artificial-Intelligence, particularly for high-stakes domains where biased predictions can have significant societal consequences. While recent advances in fair representation learning have shown promise, existing approaches often struggle with the inherent trade-off between fairness and utility, especially when addressing intersectional fairness. In this paper, we introduce Diff-Fair, a novel diffusion-based framework that leverages the progressive denoising process to learn fair representation with the help of proposed fairness constraint for intersectional bias. Our approach simultaneously addresses multiple fairness dimensions through several complementary mechanisms, (1) a diffusion model for representation extraction, (2) a mutual information estimator to minimize sensitive attribute leakage in learned representation, (3) an intersectional fairness regularizer that explicitly accounts for overlapping demographic attributes, and, (4) a false positive rate equalizer that mitigates disparate impacts across groups. Through extensive experimentation on several real-world datasets from different domain, we demonstrate that Diff-Fair consistently outperforms state-of-the-art works, reducing demographic disparities and false-positive rate difference while maintaining competitive accuracy for both binary and multi-class classification.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 13575