Explainable federated intelligence for secure and privacy-preserving distributed systems a mathematical optimization framework

Source abstract

Federated learning allows for collaborative model training without the need to share raw data, privacy protection, non-IID data, unreliable participants, and limited interpretability are all complex challenges that are intertwined. This study presents an Explainable Federated Intelligence (EFI) framework for trustworthy distributed learning, which is based on Explainability–Reliability Adaptive Privacy Federated Optimization (ERA-PF). ERA-PF assesses client updates by applying the principles of direction consistency, temporal consistency and explanation consistency. Local sample contribution is used in conjunction with these signals to compute adaptive aggregation weights. The model update is clipped prior to aggregation and reliability adaptive Gaussian perturbation is applied to the aggregated model update. Experiments use 20 heterogeneous clients, 30 communication rounds and a controlled Byzantine setting. ERA-PF had the highest accuracy (96.11%) and macro-F1 (96.07%) and the lowest expected calibration error (0.03291) of the tested configurations. The framework also yielded explicit measures of reliability and aggregation at the client level, which allowed for analysis of the influence of clients on model predictions. The results of the ablation show that the full configuration of reliability, temporal behavior, explainability and adaptive privacy is the best overall configuration. The results validate ERA-PF as a single framework for explainable, reliable and privacy-preserving federated optimization, and more adversarial testing and formal privacy accounting are needed to validate it more broadly.