Privacy-Preserving and Explainable Federated DenseNet for Mammographic Breast Cancer Classification

Abstract

Breast cancer is still a major global health problem in the world. Deep learning with the help of mammography shows great potential in early and consistent detection. However, shortcomings with patient privacy, data sharing restrictions, and limited model interpretability hinder centralized training of medical imaging models. In this study, a privacy-preserving and explainable federated DenseNet121 network approach for benign-malignant breast cancer classification is proposed with the CBIS-DDSM mammography dataset. In Flower framework, a federated learning environment is simulated with three non-independently and identically distributed (non-IID) institutions created by Dirichlet partitioning (α = 0.5), and a DenseNet121 classifier trained under federated learning settings. The aggregation strategies FedAvg, FedProx, and SCAFFOLD will be compared. At each client, Opacus is used to implement Differentially Private Stochastic Gradient Descent (DP-SGD) with privacy budgets ε ∈ {1.0, 1.9, 3.0, 5.0} to measure the privacy–utility balance. Grad-CAM/Grad-CAM++ and SHAP are used to evaluate model interpretability by identifying the regions in the image that constitute a model's predictions, and to evaluate if the model attention is affected by noise induced by privacy. Evaluation with a leak-free 5-fold cross validation at patient-level with accuracy, AUC, sensitivity and specificity as evaluation measures. McNemar’s test, Wilson score confidence intervals and Cohen’s h were added. This research discovers an appropriate balance between predictive performance and privacy, which provides clinically meaningful explanations. The proposed framework aims to provide a rigorous, privacy-aware, and interpretable approach for future multi-institutional breast cancer classification using mammography. No prior work on CBIS-DDSM has jointly compared multiple FL aggregation strategies, systematically evaluated client-side DP-SGD across a range of privacy budgets, applied Grad-CAM/Grad-CAM++ and SHAP explainability, and enforced a patient-level leak-free evaluation protocol, which constitutes the primary research gap addressed herein. The proposed SCAFFOLD-DP configuration at ε = 5.0 achieved the best classification performance with 92.12 ± 0.72% accuracy and AUC 0.940 ± 0.011, while even at the strongest privacy guarantee (ε = 1.0), SCAFFOLD retained 90.63% accuracy, representing a utility cost of approximately 6.06 percentage points relative to its non-private baseline. Regarding the effect of differential privacy on model interpretability, the mean Grad-CAM++ lesion-ROI IoU decreased progressively from 0.895 (centralized) to 0.721 (SCAFFOLD + DP, ε = 5.0) and 0.596 (SCAFFOLD + DP, ε = 1.0), indicating that stronger privacy perturbation partially reduces lesion-focused attention without eliminating clinically relevant localization. This study is subject to two principal limitations: the federated setting employs three simulated institutional clients rather than real geographically distributed hospitals, which may not fully capture real-world institutional heterogeneity; and the evaluation is confined to the CBIS-DDSM dataset, so generalization to external mammographic cohorts or other imaging modalities remains to be validated.