Are Gender-Neutral Queries Really Gender-Neutral? Mitigating Gender Bias in Image Search

TL;DR

The study shows that image search results are often gender-skewed even when queries are gender-neutral. It introduces two debiasing strategies: in-processing fair sampling for specialized image-text retrieval models and a post-processing mutual-information-based feature clipping for pre-trained multimodal representations, evaluating them on MS-COCO and Flickr30K. Results demonstrate substantial reductions in gender bias (Bias@K) with manageable tradeoffs in recall, and extended validation on internet-occupation searches. The work highlights practical methods to mitigate gender bias in multimodal search systems and lays groundwork for broader fairness in online image retrieval.

Abstract

Internet search affects people's cognition of the world, so mitigating biases in search results and learning fair models is imperative for social good. We study a unique gender bias in image search in this work: the search images are often gender-imbalanced for gender-neutral natural language queries. We diagnose two typical image search models, the specialized model trained on in-domain datasets and the generalized representation model pre-trained on massive image and text data across the internet. Both models suffer from severe gender bias. Therefore, we introduce two novel debiasing approaches: an in-processing fair sampling method to address the gender imbalance issue for training models, and a post-processing feature clipping method base on mutual information to debias multimodal representations of pre-trained models. Extensive experiments on MS-COCO and Flickr30K benchmarks show that our methods significantly reduce the gender bias in image search models.

Figures (6)

• Figure 1: Gender bias in image search. We show the top-10 retrieved images for searching "a person is cooking" on the Flickr30K Flickr30K test set using a state-of-the-art model CLIP. Despite the gender-neutral query, only 2 out of 10 images are depicting female cooking.
• Figure 2: Gender bias analysis with different top-$K$ results.
• Figure 3: The Pareto frontier of recall-bias tradeoff curve for FairSample on MS-COCO 1K and Flickr30K.
• Figure 4: Effect of the number of clipped dimensions $m$ on performance of recall and bias on MS-COCO 1K.
• Figure 5: Gender bias evaluation of internet image search results on occupationsgender-image-search. We visualize the similarity biases on 18 occupations. indicates the occupation is biased towards males and indicates it is biased towards females. The clip algorithm mitigates gender bias for a variety of occupations.