Unmasking adversarial attacks using a robust XAI-driven approach for secure medical image classification

Deep learning models have revolutionized medical image analysis but remain vulnerable to adversarial attacks. This work presents FusionStack, an explainability-driven framework for detecting adversarial perturbations through multi-method XAI stability analysis. We introduce three novel stability metrics (LIME Stability Metric, SHAP Stability Metric, Grad-CAM Stability Metric) quantifying explanation consistency across convolutional feature hierarchies, augmented with frequency-domain and gradient-based features. Tested on five medical imaging datasets, five CNN models, and five adversarial attacks with rigorous patient-level separation, FusionStack achieves 98.71% accuracy and 0.998 AUC. Cross-attack evaluation shows DeepFool-trained detectors achieve 96.85% transferability as universal defenders. Cross-dataset evaluation demonstrates 87.05% mean accuracy, with histopathology-trained detectors achieving 100% transfer to all modalities. Lightweight frequency-domain detection (10 features) achieves 95.46% accuracy with $$5.75\times$$ speedup. Our findings establish XAI stability analysis as a principled pathway toward adversarially robust medical AI systems for safety-critical applications.