BiasEdit: A Training-Free Bias-Detect-and-Edit Framework for Learning Fair Visual Classifiers

2026-05-27

AI Summary

This work addresses the challenge of unfair predictions in visual classifiers caused by implicit social biases or spurious correlations in training data, particularly under fully biased attribute distributions where learning unbiased models is especially difficult. The authors propose a training-free, modular debiasing framework that first automatically identifies unknown bias attributes through statistical dependency analysis and mutual information estimation. It then leverages vision-language representations and text-guided image editing to generate high-fidelity, debiased samples for constructing a balanced dataset. Requiring neither manual annotations, predefined bias types, nor synthetic mixing strategies, the method relies solely on off-the-shelf vision-language and image editing models. It achieves state-of-the-art debiasing performance on fully biased data, significantly enhancing both model fairness and generalization capability.

Abstract

Visual data from the Web power image classifiers, which often underpin many web services, such as recommendation and content moderation. However, the raw Web data often contain spurious correlations and social biases, and neural networks are known for their tendency to learn biases present in data. This can reinforce unfairness in web services and the web data, leading to a vicious cycle. In the context of image classification, networks learn bias attributes for a specific class when a majority of images contain the same attribute only for a given class. Hence, training a fair and debiased classifier from a biased dataset demands handling an imbalanced problem between a majority of images with bias attributes (bias-aligned samples) and a minority without (bias-conflict samples). In this work, we introduce BiasEdit, a modular framework that automatically detects bias attributes from the original dataset and edits them to construct a debiased dataset. Specifically, BiasEdit first detects unknown bias attributes via statistical dependence and mutual information analysis of visual-linguistic representations, and then explicitly edits those attributes using text-guided image editing to generate realistic bias-conflict samples. Unlike prior works that assume known bias attributes or relies on synthetic mixing, our method operates without manual annotations and can leverage off-the-shelf vision-language and editing models. BiasEdit addresses a fundamental challenge in Web-sourced visual AI, mitigating dataset-induced bias and achieving state-of-the-art debiasing performance even when training data are fully biased.

Problem

• bias
• fairness
• visual classification
• dataset bias
• spurious correlations

Innovation

• bias detection
• text-guided image editing
• training-free debiasing
• visual-linguistic representation
• fair visual classification