Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DetoxAI: a Python Toolkit for Debiasing Deep Learning Models in Computer Vision

Ignacy Stępka, Lukasz Sztukiewicz, Michał Wiliński, Jerzy Stefanowski

TL;DR

DetoxAI addresses fairness in vision-based deep learning where existing tools focus on tabular data and do not address biases in internal representations. It introduces a post-hoc, representation-level debiasing toolkit that integrates with PyTorch to desensitize protected attributes without full retraining. The framework includes multiple debiasing methods (Savani/Zhang, LEACE, ClArC variants, and Threshold Optimization), quantitative fairness metrics (Equalized Odds, Demographic Parity, Accuracy Parity), and visualization tools to track attribution shifts, all accessible via a unified detoxai.debias(...) API. The paper demonstrates DetoxAI's utility through practical engineering and research use cases, and emphasizes its production-readiness, modularity, and extensibility for broad adoption in industry and research. This work promises to streamline fairness interventions in vision models and enable robust benchmarking and deployment of debiased classifiers.

Abstract

While machine learning fairness has made significant progress in recent years, most existing solutions focus on tabular data and are poorly suited for vision-based classification tasks, which rely heavily on deep learning. To bridge this gap, we introduce DetoxAI, an open-source Python library for improving fairness in deep learning vision classifiers through post-hoc debiasing. DetoxAI implements state-of-the-art debiasing algorithms, fairness metrics, and visualization tools. It supports debiasing via interventions in internal representations and includes attribution-based visualization tools and quantitative algorithmic fairness metrics to show how bias is mitigated. This paper presents the motivation, design, and use cases of DetoxAI, demonstrating its tangible value to engineers and researchers.

DetoxAI: a Python Toolkit for Debiasing Deep Learning Models in Computer Vision

TL;DR

DetoxAI addresses fairness in vision-based deep learning where existing tools focus on tabular data and do not address biases in internal representations. It introduces a post-hoc, representation-level debiasing toolkit that integrates with PyTorch to desensitize protected attributes without full retraining. The framework includes multiple debiasing methods (Savani/Zhang, LEACE, ClArC variants, and Threshold Optimization), quantitative fairness metrics (Equalized Odds, Demographic Parity, Accuracy Parity), and visualization tools to track attribution shifts, all accessible via a unified detoxai.debias(...) API. The paper demonstrates DetoxAI's utility through practical engineering and research use cases, and emphasizes its production-readiness, modularity, and extensibility for broad adoption in industry and research. This work promises to streamline fairness interventions in vision models and enable robust benchmarking and deployment of debiased classifiers.

Abstract

While machine learning fairness has made significant progress in recent years, most existing solutions focus on tabular data and are poorly suited for vision-based classification tasks, which rely heavily on deep learning. To bridge this gap, we introduce DetoxAI, an open-source Python library for improving fairness in deep learning vision classifiers through post-hoc debiasing. DetoxAI implements state-of-the-art debiasing algorithms, fairness metrics, and visualization tools. It supports debiasing via interventions in internal representations and includes attribution-based visualization tools and quantitative algorithmic fairness metrics to show how bias is mitigated. This paper presents the motivation, design, and use cases of DetoxAI, demonstrating its tangible value to engineers and researchers.
Paper Structure (6 sections, 2 figures)

This paper contains 6 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Use Cases
  3. Engineering Use Case: Debiasing a facial expression recognition system
  4. Research Use Case: Fairness studies and comparative benchmarking
  5. System Overview
  6. Conclusions

Figures (2)

  • Figure 1: DetoxAI incorporates multiple tools to mitigate biases in vision models. Biased model with debiasing dataset is passed to our module, where biases can be inspected, measured and finally eliminated.
  • Figure 2: In the upper image, we visualize saliency maps before (Vanilla) and after bias mitigation using a selection of implemented methods. The image on the bottom is a quantitative evaluation of the debiasing techniques on the predictive performance (F1 Score) - fairness trade-off. Fairness metrics are EqualizedOdds, DemographicParity and AccuracyParity, calculated in a difference between protected groups variation mehrabi_2021_survey.