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

TL;DR

The paper tackles societal bias in image-text datasets by decorrelating protected groups from all image attributes through synthetic counterfactual data generated by text-guided inpainting. It introduces data filtering to mitigate biases arising from inpainting and demonstrates substantial bias reduction on multi-label classification and image captioning tasks without sacrificing performance. Evaluations across COCO and OpenImages with various backbones show improved leakage and bias ratio metrics, while preserving METEOR and CLIPScore-based caption quality. The work also discusses ethical considerations, limitations of real-vs-synthetic data augmentation, and directions for extending the framework to additional protected groups.

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.

Figures (9)

• Figure 1: (a) Predicted objects by baseline ResNet-50 and with bias mitigation, i.e., over-sampling wang2020towards versus our method. (b) Generated captions by baseline ClipCap and with bias mitigation, i.e., LIBRA hirota2023model versus our method. Incorrect predictions, possibly affected by gender-object correlations, are in red.
• Figure 2: Overview of our pipeline for binary gender as a protected attribute. Original images are inpainted to synthesize diverse groups, maintaining consistent context. Synthesized images (highlighted in blue) are ranked using filters to select high-quality, unbiased samples (Module: Filtering & Ranking). Selected images are then used to construct datasets with group-independent image attribute distributions (Module: Create dataset).
• Figure 3: Predicted captions for the original (left) and inpainted (right) test images.
• Figure 4: Best/worst inpainted images for each filter in \ref{['sec:method-2']} and their combination (overall).
• Figure 5: Examples of inpainted images for binary gender.