Zero-Trust Federated Learning Boosts Intrusion Detection Accuracy for Industrial IoT

Global: Zero-Trust Federated Learning Boosts Intrusion Detection Accuracy for Industrial IoT

A team of researchers has introduced Zero-Trust Agentic Federated Learning (ZTA-FL), a multi‑layered framework designed to strengthen intrusion detection in Industrial Internet of Things (IIoT) environments. The approach integrates TPM‑based cryptographic attestation, a SHAP‑weighted aggregation algorithm, and privacy‑preserving on‑device adversarial training, achieving 97.8 percent overall detection accuracy and 93.2 percent accuracy under 30 percent Byzantine attacks.

Background and Motivation

Recent compromises of critical infrastructure—including the 2021 Oldsmar water‑treatment breach and the 2023 Danish energy‑sector incidents—have highlighted persistent security gaps in IIoT deployments. These events have underscored the need for collaborative detection mechanisms that preserve privacy while resisting sophisticated adversarial tactics.

Limitations of Existing Federated Learning

Current federated learning (FL) solutions for intrusion detection often lack robust agent authentication and remain vulnerable to Byzantine poisoning attacks, especially when training data are non‑IID across participants. Such weaknesses can undermine the reliability of distributed security models.

TPM‑Based Cryptographic Attestation

ZTA‑FL incorporates Trusted Platform Module (TPM) attestation to verify the integrity of participating agents. The attestation mechanism reports a false acceptance rate of less than 0.0000001, providing a hardware‑rooted guarantee of device trustworthiness.

Explainable SHAP‑Weighted Aggregation

The framework introduces a novel SHAP‑weighted aggregation algorithm that identifies and mitigates Byzantine contributions while offering explainability. The method operates effectively under non‑IID conditions and includes theoretical guarantees for convergence and robustness.

On‑Device Adversarial Training

To preserve privacy, ZTA‑FL performs adversarial training locally on each device, reducing the exposure of raw data. This strategy enhances resilience against evasion attacks without compromising the confidentiality of sensitive industrial telemetry.

Experimental Evaluation

Comprehensive experiments were conducted on three intrusion‑detection benchmarks—Edge‑IIoTset, CIC‑IDS2017, and UNSW‑NB15. ZTA‑FL attained 97.8 percent detection accuracy overall and maintained 93.2 percent accuracy when 30 percent of participants launched Byzantine attacks, outperforming the FLAME baseline by 3.1 percent (p < 0.01). The framework also demonstrated 89.3 percent adversarial robustness and reduced communication overhead by 34 percent.

Implications and Future Work

The results suggest that integrating zero‑trust principles with federated learning can substantially improve security outcomes for IIoT systems. The authors have released code to facilitate reproducibility and plan to explore extensions for broader threat models and larger-scale deployments.

This report is based on information from arXiv, licensed under See original source. Source attribution required.