Federated Learning (FL) enables the training of deep learning models on siloed medical data. Its real-world application is often challenged by statistical heterogeneity, privacy requirements, and the need for model transparency. This paper addresses these challenges by investigating the interplay between FL, Differential Privacy (DP), and model explainability for 3D medical image segmentation. To simulate a realistic environment, we establish a cross-silo federation of four clients, comprising data from the BraTS dataset and a distinct heterogeneous dataset from a real hospital in Europe. Our analysis characterizes and quantifies an interaction, namely the phHeterogeneity Amplifier effect, providing a metric to measure the disproportionate degradation of explanation fidelity on heterogeneous clients under DP. To address this challenge, we propose Boundary-Interior Disentangled CAM (BID-CAM), a hybrid explanation method designed for DP-awareness. Our evaluation shows that BID-CAM maintains explanation fidelity under privacy constraints with respect to standard methods, demonstrating a more robust approach to model transparency in private, federated settings applied to medical imaging.