Code and artifacts for: Auditing Explanation Faithfulness in Privacy-Preserving Federated Intrusion Detection

Abstract

This archive accompanies the manuscript "Auditing Explanation Faithfulness in Privacy-Preserving Federated Intrusion Detection". It contains the source code, the seeded configuration of every run, per-run results, attribution tensors, model checkpoints, summary tables, figure code, supplementary tables and the append-only deviation log in which decision rules were recorded before the corresponding runs. The study audits the faithfulness of post-hoc explanations (Integrated Gradients, GradientSHAP, attention rollout) under client-side, example-level DP-SGD in federated, tabular intrusion detection on ToN_IoT and the corrected CICIDS2017 release, including Byzantine-robust aggregation under three attack families. The raw datasets are not redistributed; see README.md for their sources. Keywords: federated learning; differential privacy; explainable artificial intelligence; feature attribution; intrusion detection; Byzantine-robust aggregation; model poisoning