Freeze and Reveal: Exposing Modality Bias in Vision-Language Models

Vision Language Models achieve impressive multi-modal performance but often inherit gender biases from their training data. This bias might be coming from both the vision and text modalities. In this work, we dissect the contributions of vision and text backbones to these biases by applying targeted debiasing using Counterfactual Data Augmentation and Task Vector methods. Inspired by data-efficient approaches in hate-speech classification, we introduce a novel metric, Degree of Stereotypicality and a corresponding debiasing method, Data Augmentation Using Degree of Stereotypicality - DAUDoS, to reduce bias with minimal computational cost. We curate a gender annotated dataset and evaluate all methods on VisoGender benchmark to quantify improvements and identify dominant source of bias. Our results show that CDA reduces the gender gap by 6% and DAUDoS by 3% but using only one-third of the data. Both methods also improve the model's ability to correctly identify gender in images by 3%, with DAUDoS achieving this improvement using only almost one-third of training data. From our experiment's, we observed that CLIP's vision encoder is more biased whereas PaliGemma2's text encoder is more biased. By identifying whether bias stems more from vision or text encoders, our work enables more targeted and effective bias mitigation strategies in future multi-modal systems.

KTCR: Improving Implicit Hate Detection with Knowledge Transfer driven Concept Refinement

The constant shifts in social and political contexts, driven by emerging social movements and political events, lead to new forms of hate content and previously unrecognized hate patterns that machine learning models may not have captured. Some recent literature proposes the data augmentation-based techniques to enrich existing hate datasets by incorporating samples that reveal new implicit hate patterns. This approach aims to improve the model's performance on out-of-domain implicit hate instances. It is observed, that further addition of more samples for augmentation results in the decrease of the performance of the model. In this work, we propose a Knowledge Transfer-driven Concept Refinement method that distills and refines the concepts related to implicit hate samples through novel prototype alignment and concept losses, alongside data augmentation based on concept activation vectors. Experiments with several publicly available datasets show that incorporating additional implicit samples reflecting new hate patterns through concept refinement enhances the model's performance, surpassing baseline results while maintaining cross-dataset generalization capabilities.\footnote{DISCLAIMER: This paper contains explicit statements that are potentially offensive.}