ActivFairNet: A novel framework for mitigating bias in deep learning networks using activation map-based fairness regularization

The rapid advancements in Artificial Intelligence underscore the pressing need to address fairness in deep learning, particularly in critical fields like healthcare where decisions have direct and significant impacts on lives. Despite considerable progress, biases associated with sensitive attributes such as gender, race, and age remain pervasive, presenting substantial challenges to achieving equitable and reliable outcomes. This paper introduces ActivFairNet, a novel bias mitigation framework that integrates activation maps as a fairness regularizer. The framework ensures unbiased representation learning across demographic groups while maintaining or enhancing predictive accuracy, making it both effective and practical for real-world applications. The ActivFairNet is evaluated in a COVID-19 detection case study using a chest X-ray dataset collected from five public repositories. Its effectiveness was tested on three models with varying gender distributions, employing two widely recognized deep learning architectures, DenseNet121 and Xception. The results demonstrate that the ActivFairNet Regularizer consistently outperforms three established bias mitigation techniques, significantly reducing bias across key fairness metrics. Specifically, the method achieves substantial improvements, including a Statistical Parity Difference (SPD) of 0.003 (down from 0.162), an Equal Opportunity Difference (EOD) of 0.000 (down from 0.276), and an Average Odds Difference (AOD) of 0.002 (down from 0.185). The ActivFairNet Regularizer offers a practical, scalable, and ethically aligned solution for mitigating demographic bias in medical imaging, contributing to the advancement of fair and reliable AI systems in real-world healthcare environments.