Deep convolutional neural networks have achieved high accuracy in handwritten digit recognition; however, their reliability under adversarial perturbations, structured noise, and stylistic variation remains a significant challenge for real-world deployment. This paper presents a fairness-aware hybrid CNN-boosting framework that improves empirical robustness while reducing subgroup performance disparities. A convolutional neural network is employed as a feature extractor, and the resulting embeddings are classified using an ensemble of AdaBoost, XGBoost, and LightGBM models. Experiments on the EMNIST Digits dataset show that the proposed method attains 98.45% accuracy on clean data, outperforming a standalone CNN baseline (96.85%). Under Fast Gradient Sign Method (FGSM) attack with ε = 0.1, the ensemble achieves better retention stability than the baseline (0.866 vs. 0.848). The framework also demonstrates strong resilience to salt-and-pepper noise and 20% pixel occlusion. Fairness analysis across stroke-thickness subgroups indicates that loss reweighting reduces performance disparities without sacrificing overall accuracy. Cross-domain evaluation, however, reveals that distribution shift remains a persistent challenge despite gains in perturbation robustness. Overall, the results suggest that combining ensemble diversity with fairness-aware optimization offers a practical and scalable approach to building more robust and equitable handwritten digit recognition systems.
Kumar et al. (Wed,) studied this question.