README.md

metadata

license: mit
tags:
  - fairness
  - classification
metrics:
  - accuracy
papers:
  - https://arxiv.org/abs/2507.20708

Exposing the Illusion of Fairness (EIF): Biased models which results were later fairwashed

📌 Overview

This repository contains a collection of neural network models trained on seven tabular datasets for the study:

Exposing the Illusion of Fairness (EIF): Auditing Vulnerabilities to Distributional Manipulation Attacks
https://arxiv.org/abs/2507.20708

Codebase:
https://github.com/ValentinLafargue/Inspection

Results:
https://huggingface.co/datasets/ValentinLAFARGUE/EIF-Manipulated-distributions

Each model corresponds to a specific dataset and is designed to analyze fairness properties rather than maximize predictive performance.

🧠 Model Description

All models are multilayer perceptrons (MLPs) trained on tabular data.

• Fully connected neural networks
• Hidden layers: configurable (n_loop, n_nodes)
• Activation: ReLU (optional)
• Output: Sigmoid
• Prediction: $\hat{Y} \in [0,1]$

📊 Datasets, Sensitive Attributes, and Disparate Impact

Dataset	Adult[1]	INC[2]	TRA[2]	MOB[2]	BAF[3]	EMP[2]	PUC[2]
Sensitive Attribute (S)	Sex	Sex	Sex	Age	Age	Disability	Disability
Disparate Impact (DI)	0.30	0.67	0.69	0.45	0.35	0.30	0.32

[1]: Becker, B. and Kohavi, R. (1996). Adult. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5XW20.306,
https://www.kaggle.com/datasets/uciml/adult-census-income.

[2]: Ding, F., Hardt, M., Miller, J., and Schmidt, L. (2021). Retiring adult: New datasets for fair machine learning. In Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W., editors, Advances in Neural Information Processing Systems.313,
https://github.com/socialfoundations/folktables.

[3]: Jesus, S., Pombal, J., Alves, D., Cruz, A., Saleiro, P., Ribeiro, R. P., Gama, J., and Bizarro, P. (2022). Turning the tables: Biased, imbalanced, dynamic tabular datasets for ml evaluation. In Advances in Neural Information Processing Systems,
https://www.kaggle.com/datasets/sgpjesus/bank-account-fraud-dataset-neurips-2022.

Notes

• Adult dataset: 5,000 test samples
• Other datasets: 20,000 test samples
• Sensitive attributes are used for fairness evaluation

Results and manipulated results

The results obtained on the tests samples, and their fairwashed counterparts are directly available on Hugging Face.

📈 Predictive Performance (Accuracy)

Dataset	Accuracy
Adult Census Income	84%
Folktables Income (INC)	88%
Folktables Mobility (MOB)	84%
Folktables Employment (EMP)	77%
Folktables Travel Time (TRA)	72%
Folktables Public Coverage (PUC)	73%
Bank Account Fraud (BAF)	98%

Note: High performance on BAF is due to strong class imbalance.
Accuracy was not the main objective of this study.

🎯 Intended Use

These models are intended for:

• Fairness analysis
• Studying disparate impact and bias
• Reproducing results from the EIF paper
• Benchmarking fairness-aware methods

⚠️ Limitations and Non-Intended Use

• Not designed for production
• Not optimized for predictive performance
• Should not be used for real-world decision-making

These models intentionally expose biases in standard ML pipelines.

⚖️ Ethical Considerations

This work highlights:

• The presence of bias in machine learning models
• The limitations of fairness metrics

Models should be interpreted as analytical tools, not fair systems.

📦 Repository Structure

Each dataset corresponds to a subfolder:

EIF-biased-classifier/
├── ASC_ADULT_model/
├── ASC_INC_model/
├── ASC_MOB_model/
├── ASC_EMP_model/
├── ASC_TRA_model/
├── ASC_PUC_model/
└── ASC_BAF_model/

Each folder contains:

• config.json
• model.safetensors

🚀 Usage

model = Network.from_pretrained(
    "ValentinLAFARGUE/EIF-biased-classifier",
    subfolder="ASC_INC_model"
)

📚 Citation

@misc{lafargue2026exposingillusionfairnessauditing,
      title={Exposing the Illusion of Fairness: Auditing Vulnerabilities to Distributional Manipulation Attacks}, 
      author={Valentin Lafargue and Adriana Laurindo Monteiro and Emmanuelle Claeys and Laurent Risser and Jean-Michel Loubes},
      year={2026},
      eprint={2507.20708},
      url={https://arxiv.org/abs/2507.20708}, 
}

🔍 Additional Notes

• Models are intentionally simple to isolate fairness behavior
• Results depend on preprocessing and sampling choices
• Focus is on reproducibility