Visual Explanations from Deep Networks Using GradCAM in Federated Learning

In the realm of Explainable Artificial Intelligence (XAI), visual explanations play a crucial role in demystifying complex models, providing insights into their decision-making processes. This paper explores the integration of Federated Learning (FL) as a method to enhance privacy without compromising the quality of visual explanations. FL, by design, allows multiple decentralized devices to collaboratively train models while keeping the data localized, thereby significantly mitigating privacy concerns. We specifically evaluate the impact of FL on visual explanations generated by Gradient-weighted Class Activation Mapping (Grad-CAM), a popular technique for producing class-discriminative localization maps. Our study demonstrates that while FL introduces certain challenges, such as heterogeneity of data across different devices, it does not substantially degrade the interpretability of models as visualized by Grad-CAM. This research highlights the feasibility and benefits of combining FL with visual explanation techniques, paving the way for more secure and interpretable AI systems.