[논문 리뷰] Fairness via Representation Neutralization

Mengnan Du, Subhabrata Mukherjee외 4명|arXiv (Cornell University)|2021. 06. 23.

Ethics and Social Impacts of AI참고 문헌 42인용 수 28

한 줄 요약

RNF은 중립화된 표현을 사용하여 분류 헤드만 편향 제거함으로써 편향을 완화하고, 필요하면 프록시 민감 속성 주석을 사용하며, 학습 중 보호 속성 라벨이 필요하지 않다.

ABSTRACT

Existing bias mitigation methods for DNN models primarily work on learning debiased encoders. This process not only requires a lot of instance-level annotations for sensitive attributes, it also does not guarantee that all fairness sensitive information has been removed from the encoder. To address these limitations, we explore the following research question: Can we reduce the discrimination of DNN models by only debiasing the classification head, even with biased representations as inputs? To this end, we propose a new mitigation technique, namely, Representation Neutralization for Fairness (RNF) that achieves fairness by debiasing only the task-specific classification head of DNN models. To this end, we leverage samples with the same ground-truth label but different sensitive attributes, and use their neutralized representations to train the classification head of the DNN model. The key idea of RNF is to discourage the classification head from capturing spurious correlation between fairness sensitive information in encoder representations with specific class labels. To address low-resource settings with no access to sensitive attribute annotations, we leverage a bias-amplified model to generate proxy annotations for sensitive attributes. Experimental results over several benchmark datasets demonstrate our RNF framework to effectively reduce discrimination of DNN models with minimal degradation in task-specific performance.

연구 동기 및 목표

• Motivate and analyze how bias propagates from encoder representations to the task-specific classification head.
• Propose Representation Neutralization for Fairness (RNF) to debias only the classification head while using biased encoders as input.
• Enable fairness mitigation in low-resource settings by generating proxy annotations for sensitive attributes without access to ground-truth labels.
• Demonstrate that RNF reduces discrimination across tabular and image benchmarks with minimal loss in task accuracy.
• Show that RNF complements existing debiased-encoder methods and can be combined with them for improved fairness.

제안 방법

• Train a biased teacher network with standard cross entropy to obtain biased representations.
• Freeze the encoder from the biased teacher and re-train only the classification head using representation neutralization, which uses the average of two representations from samples with the same label but different sensitive attributes.
• Use softened probabilities from the biased teacher (temperature scaling) as targets for knowledge distillation during head retraining.
• Incorporate a smoothing neutralization and a mixed loss (MSE between neutralized representations and softened predictions) to encourage reliance on task-relevant information and reduce sensitivity to protected attributes.
• Optionally generate proxy sensitive-attribute annotations via a bias-amplified model trained with generalized cross entropy (GCE) to enable RNF in settings without attribute labels.
• Combine the biased encoder with the retrained head to form a debiased student model for inference.

실험 결과

연구 질문

• RQ1 Can debiasing the classification head alone, given biased encoder representations, reduce discrimination without heavily impacting task performance?
• RQ2 How can proxy annotations for sensitive attributes be generated when protected attribute labels are unavailable, and how effective are they for RNF?
• RQ3 What is the effect of neutralizing representations on the reliance of the head on fairness-sensitive information?
• RQ4 How does RNF compare to adversarial training and equalized odds regularization in terms of fairness-accuracy trade-offs?
• RQ5 Can RNF complement existing debiasing methods by serving as an add-on to debiased encoders?

주요 결과

• RNF significantly reduces discrimination across benchmark datasets with minimal task-accuracy degradation.
• RNF performs comparably to methods that use ground-truth sensitive attributes and can outperform some baselines in certain metrics.
• Using softened labels (temperature T > 1) and the smoothing regularization improves mitigation by discouraging overconfident, biased predictions.
• GCE-based proxy annotations enable RNF in low-resource settings, achieving meaningful fairness improvements even with imperfect annotations.
• Debiasing only the last layer (RNF-Last) can yield substantial fairness gains with small or comparable accuracy loss to debiasing deeper layers.
• RNF can complement debiased encoders; combining RNF with known debiasing encoders further improves fairness metrics.