Work

RPP-FL: A Robust and Privacy-aware Federated Learning Framework for IIoT Anomaly Detection under Non-IID Data and Poisoning Attacks

Year: 2026
Type: article
Abstract: Federated learning (FL) enables collaborative intrusion detection without centralizing raw Industrial Internet of Things (IIoT) traffic, yet its practical reliability is challenged by non-IID client data, poisoning attacks, and privacy leakage through shared updates.This paper proposes RPP-FL, a robust and privacy-aware federated learning framework that combines a compact feature-token Transformer-MLP detector with multi-criterion server-side update screening.RPP-FL integrates robust coordinate deviation, center and momentum consistency, peer agreement, compact-minority collusion detection, persistent asymmetric reputation, reputation-aware weighting, coordinate-wise winsorization, and adaptive median anchoring.The methodological contribution lies in the event-coupled propagation of compact-minority collusion evidence across instantaneous trust, persistent client reputation, and adaptive medianbased aggregation.Rather than introducing these individual mechanisms as new primitives, RPP-FL investigates their coordinated operation and the resulting security-utility trade-offs under heterogeneous and adversarial federatedlearning conditions.Privacy is evaluated in a separate aggregate-release mode using uniformly weighted, individually clipped client updates and client-level Gaussian differential privacy with Rényi accounting.This privacy experiment is operationally distinct from the RPP robust-screening mode, which requires visibility of individual client updates.The present work therefore does not claim simultaneous Byzantine screening and cryptographic secure aggregation, nor does it implement or benchmark a concrete secure-aggregation protocol.Experiments on CICIoT2023, EdgeIIoTset, TON_IoT, and WUSTL-IIoT yield a mean clean accuracy of 99.605%, F1-score of 99.739%, and ROC-AUC of 99.911%.Under 20% ALIE poisoning, mean F1 remains 99.671%, corresponding to an average reduction of only 0.068 percentage points.An additional three-seed mechanism-level evaluation on TON_IoT showed that full event coupling reduced malicious aggregation weight from 25.161% with simpler temporal-trust aggregation to approximately zero under ALIE-20% poisoning.However, partial coupling achieved comparable suppression, while P4P retained higher predictive F1.These findings demonstrate condition-dependent benefits rather than consistent superiority of the complete coupling mechanism.High-malicious-ratio and targeted adaptive stress tests further showed that high end-task F1 does not necessarily imply reliable malicious-client attribution, particularly under whitebox trust-evasion attacks.Under severe non-IID partitioning (Dirichlet α = 0.1), mean F1-scores reached 99.530% and 99.594% on CICIoT2023 and 98.457% and 98.502% on TON_IoT under clean and ALIE-20% conditions, respectively, highlighting dataset-dependent performance under heterogeneous client distributions.Taken together, these results demonstrate consistent benchmark-specific detection performance across multiple independently trained IoT/IIoT datasets, together with resilience to heterogeneous client partitions and the evaluated adversarial conditions.Because each dataset was trained and evaluated independently, these experiments should not be interpreted as evidence of cross-dataset transfer or zero-shot generalization.
Language: English
Open Access status: hybrid
OpenAlex ID: W7234900413
FWCI:
Cites: 35
Cited by:
Related to: 10