[Paper Review] FairNeuron: Improving Deep Neural Network Fairness with Adversary Games on Selective Neurons

Xuanqi Gao, Juan Zhai+4 more|arXiv (Cornell University)|Apr 6, 2022

Adversarial Robustness in Machine Learning4 citations

TL;DR

FairNeuron is a lightweight, post-hoc method that improves deep neural network fairness by identifying and selectively dropping neurons in conflict paths—those that over-rely on sensitive attributes—using path analysis and selective retraining. It effectively balances fairness and accuracy without retraining the entire model, outperforming state-of-the-art methods in efficiency and scalability on multiple datasets.

ABSTRACT

With Deep Neural Network (DNN) being integrated into a growing number of critical systems with far-reaching impacts on society, there are increasing concerns on their ethical performance, such as fairness. Unfortunately, model fairness and accuracy in many cases are contradictory goals to optimize. To solve this issue, there has been a number of work trying to improve model fairness by using an adversarial game in model level. This approach introduces an adversary that evaluates the fairness of a model besides its prediction accuracy on the main task, and performs joint-optimization to achieve a balanced result. In this paper, we noticed that when performing backward propagation based training, such contradictory phenomenon has shown on individual neuron level. Based on this observation, we propose FairNeuron, a DNN model automatic repairing tool, to mitigate fairness concerns and balance the accuracy-fairness trade-off without introducing another model. It works on detecting neurons with contradictory optimization directions from accuracy and fairness training goals, and achieving a trade-off by selective dropout. Comparing with state-of-the-art methods, our approach is lightweight, making it scalable and more efficient. Our evaluation on 3 datasets shows that FairNeuron can effectively improve all models' fairness while maintaining a stable utility.

Motivation & Objective

• To address the trade-off between fairness and accuracy in deep neural networks, especially when model fairness and utility are conflicting goals.
• To detect neurons and paths that disproportionately rely on sensitive attributes (e.g., race, gender) for predictions, leading to unfairness.
• To develop a lightweight, efficient, and scalable method for repairing fairness issues without retraining the full model or introducing auxiliary networks.
• To enable automatic, post-training fairness repair through path analysis and selective dropout, preserving model utility.

Proposed method

• Uses neural network slicing to identify conflict paths—subnetworks where neurons selectively use sensitive attributes over distinguishable features.
• Applies path analysis to cluster input samples based on their activation of conflict paths, identifying bias-triggering inputs.
• Performs selective retraining by dropping neurons outside conflict paths during backpropagation, forcing the model to reconsider non-sensitive features.
• Leverages gradient-based analysis to detect neurons with conflicting optimization directions between accuracy and fairness objectives.
• Introduces a lightweight, model-agnostic repair procedure that operates after training, avoiding complex adversarial training or data augmentation.
• Employs a dual-objective optimization: maintaining prediction accuracy while minimizing unfairness via targeted neuron suppression.

Experimental results

Research questions

• RQ1 Can we detect specific neurons and paths in a DNN that contribute to unfair predictions by over-relying on sensitive attributes?
• RQ2 How can we improve fairness without degrading model accuracy or requiring full retraining?
• RQ3 Can selective neuron dropout based on path analysis effectively mitigate fairness issues in a lightweight and scalable manner?
• RQ4 How does FairNeuron compare to state-of-the-art fairness repair methods in terms of fairness gain and efficiency?

Key findings

• FairNeuron significantly improves fairness across all evaluated models on three benchmark datasets without degrading prediction accuracy.
• The method achieves better fairness improvements than state-of-the-art approaches like FAD and Ethical Adversaries, particularly in terms of efficiency and scalability.
• Selective retraining via conflict path analysis reduces reliance on sensitive attributes, especially in fully connected layers, where fairness gains are most needed.
• The approach is lightweight and model-agnostic, making it suitable for large-scale DNNs where adversarial training or data augmentation is impractical.
• Evaluation shows stable utility across fairness metrics, including demographic parity and equal opportunity, across different sensitive attribute groups.