Dual-level Bias Mitigation via Fairness-guided Distribution Discrepancy
Keywords: Fairness, Trust-worthy Machine Learning
TL;DR: We propose a method to ensure fairness in fine-tuned models by optimizing fairness-guided discrepancies, backed by strong theoretical guarantees.
Abstract: Modern artificial intelligence predominantly relies on pre-trained models, which are fine-tuned for specific downstream tasks rather than built from scratch. However, a key challenge persists: the fairness of learned representations in pre-trained models is not guaranteed when transferred to new tasks, potentially leading to biased outcomes, even if fairness constraints were applied during the original training. To address this issue, we propose Dual-level Bias Mitigation (DBM), which measures the fairness-guided distribution discrepancy between representations of different demographic groups. By optimizing both the fairness-guided distribution discrepancy and the task-specific objective, DBM ensures fairness at both the representation and task levels. Theoretically, we provide the generalization error bound of the fairness-guided distribution discrepancy to support the efficacy of our approach. Experimental results on multiple benchmark datasets demonstrate that DBM effectively mitigates bias in fine-tuned models on downstream tasks across a range of fairness metrics.
Supplementary Material: zip
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8438
Loading