DetoxAI Logo DetoxAI

A Python package for debiasing neural networks

Ignacy Stepka* Lukasz Sztukiewicz* Michal Wilinski* Jerzy Stefanowski
PUT Logo Poznan University of Technology
*Indicates Equal Contribution
Published at ECML PKDD 2025, Demo Track
Code Documentation Video Paper Springer

How DetoxAI Works

DetoxAI Features
DetoxAI transforms biased neural networks into fair and balanced models with minimal code integration.

Key Features

Multiple Debiasing Methods

DetoxAI includes several state-of-the-art debiasing techniques:

  • ClArC family
  • Zhang et al.
  • Savani et al.
  • Belrose et al.
Learn how to add your own method →

Dataset Integration

Works with popular datasets out of the box:

  • CelebA
  • FairFace
  • CIFAR-10/100
  • Caltech101
Learn how to use datasets →

Visualization Tools

Powerful tools to understand and visualize bias:

  • Saliency maps with Layer-wise Relevance Propagation
  • Side-by-side comparisons of original vs. debiased models
  • Aggregate visualizations to track focus shift
Explore visualization tools →

Simple API

DetoxAI is designed to be easy to use:

  • Works with existing PyTorch models
  • Integrates with standard dataloaders
  • Single function call to apply multiple methods

Quick Start

Code Example 
# DetoxAI is available on PyPI, and can be installed by running the following command:
pip install detoxai

# DetoxAI Quickstart Example
import detoxai

# Load your pretrained model
model = ...

# Prepare a dataloader yielding (input, label, protected attribute)
dataloader = ...

# Debias the model using DetoxAI
corrected = detoxai.debias(model, dataloader)

# Access metrics from the SAVANIAFT debiasing method
metrics = corrected["SAVANIAFT"].get_all_metrics()

# Retrieve the debiased model
model = corrected["SAVANIAFT"].get_model()

DetoxAI provides a simple API for debiasing neural networks in image classification tasks.

Side by side comparison

Saliency maps obtained with Layer-wise Relevance Propagation (LRP) showing where the model focuses when classifying images.

Aggregate results

Aggregate visualizations (average photos) that we used in our work to show the effect of debiasing.

BibTeX


          @inproceedings{stkepka2025detoxai,
            title={DetoxAI: a Python Toolkit for Debiasing Deep Learning Models in Computer Vision},
            author={St{\k{e}}pka, Ignacy and Sztukiewicz, Lukasz and Wili{\'n}ski, Micha{\l} and Stefanowski, Jerzy},
            booktitle={Joint European Conference on Machine Learning and Knowledge Discovery in Databases},
            pages={502--505},
            year={2025},
            organization={Springer}
          }