FedStaP: code, trained models and results for non-IID federated intrusion detection (CICIoT2023, Edge-IIoTset)

Abstract

Research artefact for the study FedStaP: Shared Feature Statistics and Prior-Calibrated Training for Non-IID Federated Intrusion Detection in IoT Networks. Contents. The full implementation (Hydra configurations, federated simulator, the proposed FedStaP method, ten global-model baselines, a personalised reference and a centralised reference, and eight ablation variants); the validation-selected global model of each of the 738 experiment runs, with that run's configuration and test metrics; all per-run results, learning curves, analysis scripts, LaTeX tables and figures; an independent re-derivation script that reproduces every number reported in the manuscript from the raw run records; and a locked dependency list. Protocol. Two public datasets (CICIoT2023 and Edge-IIoTset), two label granularities, Dirichlet label-skew partitions, five seeds for the main comparison, a full 2^3 ablation, sensitivity sweeps over client count, skew, participation and calibration strength, and Wilcoxon/Holm and Friedman/Nemenyi statistical testing. Datasets. CICIoT2023 and Edge-IIoTset are not redistributed here. The preparation scripts and the exact preparation metadata (source hashes, feature lists, class counts, per-class caps and sampling seed) are included, so the prepared data can be regenerated from the public sources.