On the Adaptability of ML Reliability Estimation Techniques to Federated Settings

Federated learning (FL) enables collaborative model training without centralizing data, thereby preserving privacy, but it introduces new challenges for assessing the reliability of AI predictions. Classical reliability estimation methods from centralized machine learning cannot be directly applied in FL due to data heterogeneity, limited observability, and communication constraints. This paper presents an early-stage investigation into adapting representative reliability techniques-spanning data-driven, model-driven, and complementary information-based methods-to the federated setting. For each category, we analyze feasibility, outline privacy-preserving adaptations, and discuss trade-offs in terms of accuracy, computation, and communication. While no experimental validation is provided at this stage, the study offers preliminary insights that may guide future research and empirical evaluation of reliable FL systems.