PhysGuard-FL: Reproducibility package for physics-aware Byzantine-robust federated learning in chemical process monitoring

Abstract

Complete research compendium for PhysGuard-FL, a physics-aware Byzantine-robust federated-learning framework evaluated for chemical-process monitoring. The frozen benchmark uses six Tennessee Eastman Process operating modes, 20 federated clients, 20 communication rounds, five local steps, and 10 matched random seeds. The primary robustness benchmark comprises 70 runs across seven aggregation methods under model-replacement attacks with 30% Byzantine clients. The secondary benchmark comprises 40 runs across four shared aggregation methods under sign-flip attacks with 20% Byzantine clients. The archive preserves all 110 run-level experiment outputs, original summaries, Student-t 95% confidence intervals, paired t-tests, Wilcoxon signed-rank tests, Holm-adjusted p-values, seed-wise wins and ranks, cross-attack comparisons, figures, documentation, and executable statistical post-processing scripts. The package supports transparent statistical auditing and reproducibility of the quantitative results reported in the associated PhysGuard-FL study.