Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA

Songqi Zhou, Zeyuan Liu, Benben Jiang

TL;DR

FairNet proposes a dynamic, in-model fairness correction by integrating a bias detector with conditional LoRA adapters, activated on biased instances to preserve overall accuracy. A novel contrastive loss aligns minority and majority representations within the same class, addressing minority underfitting and reducing disparities in metrics like worst-group accuracy and equalized odds. The framework provides theoretical guarantees under reasonable detector performance and demonstrates strong empirical results across vision and language benchmarks in fully labeled, partially labeled, and unlabeled attribute scenarios. Practically, FairNet offers a scalable, modular approach to fair AI that maintains high task performance while improving outcomes for underrepresented groups.

Abstract

Ensuring fairness in machine learning models is a critical challenge. Existing debiasing methods often compromise performance, rely on static correction strategies, and struggle with data sparsity, particularly within minority groups. Furthermore, their utilization of sensitive attributes is often suboptimal, either depending excessively on complete attribute labeling or disregarding these attributes entirely. To overcome these limitations, we propose FairNet, a novel framework for dynamic, instance-level fairness correction. FairNet integrates a bias detector with conditional low-rank adaptation (LoRA), which enables selective activation of the fairness correction mechanism exclusively for instances identified as biased, and thereby preserve performance on unbiased instances. A key contribution is a new contrastive loss function for training the LoRA module, specifically designed to minimize intra-class representation disparities across different sensitive groups and effectively address underfitting in minority groups. The FairNet framework can flexibly handle scenarios with complete, partial, or entirely absent sensitive attribute labels. Theoretical analysis confirms that, under moderate TPR/FPR for the bias detector, FairNet can enhance the performance of the worst group without diminishing overall model performance, and potentially yield slight performance improvements. Comprehensive empirical evaluations across diverse vision and language benchmarks validate the effectiveness of FairNet.

FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA

TL;DR

FairNet proposes a dynamic, in-model fairness correction by integrating a bias detector with conditional LoRA adapters, activated on biased instances to preserve overall accuracy. A novel contrastive loss aligns minority and majority representations within the same class, addressing minority underfitting and reducing disparities in metrics like worst-group accuracy and equalized odds. The framework provides theoretical guarantees under reasonable detector performance and demonstrates strong empirical results across vision and language benchmarks in fully labeled, partially labeled, and unlabeled attribute scenarios. Practically, FairNet offers a scalable, modular approach to fair AI that maintains high task performance while improving outcomes for underrepresented groups.

Abstract

Ensuring fairness in machine learning models is a critical challenge. Existing debiasing methods often compromise performance, rely on static correction strategies, and struggle with data sparsity, particularly within minority groups. Furthermore, their utilization of sensitive attributes is often suboptimal, either depending excessively on complete attribute labeling or disregarding these attributes entirely. To overcome these limitations, we propose FairNet, a novel framework for dynamic, instance-level fairness correction. FairNet integrates a bias detector with conditional low-rank adaptation (LoRA), which enables selective activation of the fairness correction mechanism exclusively for instances identified as biased, and thereby preserve performance on unbiased instances. A key contribution is a new contrastive loss function for training the LoRA module, specifically designed to minimize intra-class representation disparities across different sensitive groups and effectively address underfitting in minority groups. The FairNet framework can flexibly handle scenarios with complete, partial, or entirely absent sensitive attribute labels. Theoretical analysis confirms that, under moderate TPR/FPR for the bias detector, FairNet can enhance the performance of the worst group without diminishing overall model performance, and potentially yield slight performance improvements. Comprehensive empirical evaluations across diverse vision and language benchmarks validate the effectiveness of FairNet.
Paper Structure (42 sections, 19 equations, 2 figures, 10 tables)

This paper contains 42 sections, 19 equations, 2 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Diversity and Challenges of Fairness Definitions and Metrics
  4. Fairness Intervention Strategies and Performance-Fairness Trade-off
  5. Fairness under Varying Sensitive Attribute Constraints
  6. Method
  7. Framework Overview and Inference Process
  8. Bias Detection Module
  9. Contrastive Conditional LoRA Module
  10. Model Training
  11. Theoretical Analysis
  12. Preliminaries
  13. Contrastive Loss and Representation Fairness
  14. Performance Preservation Analysis
  15. Experiments
  16. ...and 27 more sections

Figures (2)

  • Figure 1: FairNet framework and conceptual training steps. (Top) The overall architecture featuring Bias Detectors and Contrastive Conditional LoRA (CL) modules. (Bottom) A four-step conceptual training process: (1) Base model preparation, (2) Bias Detector training, (3) Contrastive pair embedding preparation, and (4) Conditionally training LoRA modules with Contrastive Loss based on detector signals.
  • Figure 2: TPR and FPR Analysis with TPR/FPR Ratio for African American and Female Groups across Different Thresholds on the HateXplain dataset.