Federated Explainable AI: A Comprehensive Review and Taxonomy

This comprehensive survey delves into the emerging field of Federated Explainable Artificial Intelligence (FedXAI), which combines the privacy-preserving benefits of Federated Learning (FL) with the transparency and trust offered by Explainable AI (XAI). While FL effectively addresses data confidentiality by keeping raw data local, it doesn't inherently resolve the inherent opacity of modern machine learning models. FedXAI aims to bridge this gap, ensuring both privacy and explainability are met. The paper emphasizes how explainability is transitioning from a post-hoc analysis tool to an integral component throughout the entire FL lifecycle, supporting critical functions such as aggregation, personalization, robustness, and system-level decision-making. To organize the diverse literature, the authors propose a new taxonomy that categorizes FedXAI methods based on the role of explainability, types of models and explainers, explanation scope, integration level, FL settings, and data heterogeneity. The review covers a wide range of approaches, from model-agnostic explanations to inherently interpretable federated models and aggregation mechanisms that are aware of explainability. The survey also critically examines current evaluation practices, pointing out a significant lack of standardized benchmarks and metrics for assessing explanation quality, stability, potential privacy leakage, and computational overhead in FedXAI systems. Finally, it identifies several key open challenges that need to be addressed, including achieving explainability under non-IID (non-independent and identically distributed) data conditions, mitigating explanation-centric security threats, developing communication-efficient XAI, enabling continual FedXAI, and effectively integrating domain knowledge and regulatory constraints into the framework.