Resampled Datasets Are Not Enough: Mitigating Societal Bias Beyond Single Attributes

TL;DR

This paper presents a novel approach to reduce societal bias in image-text datasets by removing spurious correlations across all attributes using text-guided inpainting, improving fairness without sacrificing model performance.

Contribution

It introduces a method that mitigates biases from both labeled and unlabeled attributes, surpassing traditional single-attribute bias removal techniques.

Findings

Effectively reduces societal bias in image-text datasets.

Maintains performance across image classification and captioning tasks.

Uses text-guided inpainting for comprehensive bias mitigation.

Abstract

We tackle societal bias in image-text datasets by removing spurious correlations between protected groups and image attributes. Traditional methods only target labeled attributes, ignoring biases from unlabeled ones. Using text-guided inpainting models, our approach ensures protected group independence from all attributes and mitigates inpainting biases through data filtering. Evaluations on multi-label image classification and image captioning tasks show our method effectively reduces bias without compromising performance across various models.