Responsible Artificial Intelligence Framework For Fair, Explainable, And Privacy-Aware Decision Systems

Abstract

There is a growing number of deep learning decision systems in sensitive areas, however some of these systems suffer from algorithmic biases, lack of interpretability and privacy concerns. The current ways of tackling fairness, explainability, and privacy are isolated and fail to offer a comprehensive responsible AI solution. This study presents an integrated Responsible Artificial Intelligence approach which consists of a Random Forest decision model, reweighing based fairness mitigation, SHAP based explanation generation and Laplace differential privacy. The framework was tested via a classification experiment simulated with respect to accuracy, F1-score, demographic parity difference, equal opportunity difference, explanation fidelity, privacy budget and processing time. The proposed framework has a % accuracy of 94.2 and F1 score of 93.8. It achieved an explanation fidelity of 96.1%, and decreased the demographic parity difference as well as the equal opportunity difference from 0.214 to 0.061 and from 0.187 to 0.052 respectively. Even when ε = 10% (privacy budget), the accuracy level drops just a 1.8 percentage points along with an average decision time of 18.6ms. The results suggest that the model has a good performance in terms of its predictiveness, fairness, explainability and privacy.