CI+KL: Confidence and divergence-guided aggregation with mixed-precision training for robust federated learning

Abstract

Federated learning (FL) performance often degrades under non-independent and non-identically distributed (non-IID) client data, noisy updates, and unstable optimization dynamics. In this work, we propose CI+KL, a federated aggregation method that combines two complementary weighting signals: confidence interval (CI) widths, which reflect the statistical reliability of client updates, and Kullback–Leibler (KL) divergence, which captures distributional alignment with a global reference. By jointly accounting for update uncertainty and distributional mismatch, CI+KL adaptively modulates client contributions during aggregation. To improve computational efficiency, CI+KL is integrated with mixed-precision training, using FP16 for local computation and FP32 for global aggregation, reducing memory usage while preserving model accuracy. Experiments on standard benchmarks (CIFAR-10, CIFAR-100, SVHN, MNIST, and Shakespeare) demonstrate that CI+KL achieves stable convergence and competitive accuracy compared to established FL baselines under heterogeneous data distributions. Theoretical analysis supports the use of CIwidth as a proxy for variance, motivating the variance-reduction effect observed empirically. Overall, CI+KL provides a statistically grounded aggregation framework that demonstrates robustness and efficiency improvements on standard federated learning benchmarks. © 2026 Elsevier Inc.