Leveraging Statistical Invariants to Fortify CNN-Based Medical Diagnostics Against Adversarial Perturbations

The integration of artificial intelligence (AI) in medical diagnostics is increasingly jeopardized by adversarial attacks—imperceptible perturbations designed to induce misclassification in Deep Learning models. While Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in medical image analysis, their susceptibility to gradient-based attacks poses a severe risk to patient safety and diagnostic integrity. This study addresses the critical need for robust defense mechanisms in X-ray diagnostics by proposing a Hybrid Ensemble model based on Stacked Generalization. Unlike single-paradigm approaches, our method fuses the spatial feature extraction capabilities of a CNN with the statistical anomaly detection power of a Random Forest (RF). We evaluated this architecture on a curated dataset of X-ray images subjected to Projected Gradient Descent (PGD) attacks with varying perturbation magnitudes (ϵ). The results demonstrate that the Hybrid Ensemble consistently outperforms individual models and standard adversarial training baselines. Under strong attack conditions (ϵ = 0.006), the proposed model achieved an Area Under the Curve (AUC) of 0.919, significantly surpassing the adversarial training baseline (AUC 0.700). Furthermore, the ensemble reduced false positives to 108 compared to 138 for the CNN alone, enhancing clinical reliability. Theoretical motivation for the feature extraction process and extensive experimental validation suggest that leveraging statistical irregularities offers a computationally efficient and robust defense strategy suitable for real-time clinical deployment.