BiasDora: Exploring Hidden Biased Associations in Vision-Language Models

TL;DR

BiasDora investigates hidden implicit biases in vision-language models across multiple bias dimensions, revealing subtle and extreme biases often overlooked by existing methods, and provides a dataset for further research.

Contribution

The paper introduces a systematic approach to uncover hidden implicit biases in VLMs across 9 bias dimensions, expanding beyond documented associations.

Findings

Identifies subtle and extreme biases in VLMs

Reveals variation in negativity, toxicity, and extremity of biases

Provides a publicly available dataset of bias associations

Abstract

Existing works examining Vision-Language Models (VLMs) for social biases predominantly focus on a limited set of documented bias associations, such as gender:profession or race:crime. This narrow scope often overlooks a vast range of unexamined implicit associations, restricting the identification and, hence, mitigation of such biases. We address this gap by probing VLMs to (1) uncover hidden, implicit associations across 9 bias dimensions. We systematically explore diverse input and output modalities and (2) demonstrate how biased associations vary in their negativity, toxicity, and extremity. Our work (3) identifies subtle and extreme biases that are typically not recognized by existing methodologies. We make the Dataset of retrieved associations, (Dora), publicly available here https://github.com/chahatraj/BiasDora.