A Multi-Seed Evaluation of Explanation Faithfulness and Stability under Adversarial Evasion in IoT/IIoT Intrusion Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper presents a controlled, multi-seed study of adversarial evasion and explanation quality for tabular IoT/IIoT intrusion detection. ToN-IoT is used as the primary reproduction benchmark and Edge-IIoTset as an external validation benchmark. The experiment fixes the binary task, feature preprocessing, partitions, MLP architecture, five training seeds, constrained PGD/FGSM attacks, and SHAP/LIME evaluation protocol. Clean and adversarially trained controls are evaluated using predictive, evasion, faithfulness, stability, runtime, and seed-level statistical metrics. On the completed run, clean ToN-IoT models achieve mean macro-F1 0.9697 but their PGD attack success reaches 0.8073 at ϵ = 0.05; adversarial training reduces this to 0.6134 while lowering clean macro-F1 to 0.8658. On Edge-IIoTset, adversarial training reduces PGD success from 0.3511 to 0.0989 and increases adversarial malicious recall from 0.5033 to 0.6453. Explanation behavior is dataset- and explainer-dependent: SHAP is approximately twelve times slower than LIME, ToN-IoT top-10 clean–adversarial overlap is 1.0, whereas Edge-IIoTset overlap ranges from approximately 0.51 to 0.75. Seed-level differences are reported through Cohen’s dz and seed-bootstrap confidence intervals; with five seeds, no SHAP–LIME comparison crosses the Holm-adjusted α = 0.05 threshold. The results close a methodological gap by jointly measuring detection, evasion, faithfulness, stability, and measured run-time under a reproducible multi-seed protocol, while retaining cautious claims about the bounded feature-space threat model.