Multi-AI Agent Oriented Privacy Policy Compliance Checking Based on Large Language Models in Mobile IoT Systems

Zeng, Yangyan, Liu, Wei, Wang, Anqi, Zeng, Chunchao, Yang, Yi, Liang, Wei, Zhang, Xu ORCID: https://orcid.org/0000-0001-6557-6607, Jamal Deen, M. and Zhou, Xiaokang (2026) Multi-AI Agent Oriented Privacy Policy Compliance Checking Based on Large Language Models in Mobile IoT Systems. IEEE Internet of Things Journal, 13 (18). pp. 41456-41502. ISSN 2327-4662

Preview

PDF (Multi-AI Agent) - Accepted Version Available under License Creative Commons Attribution. Download (1MB) | Preview

Abstract

The proliferation of the Internet of Things (IoT) has exacerbated data privacy risks, necessitating rigorous compliance auditing of IoT companion application privacy policies. However, automated detection faces unique technical challenges, such as complex cross-paragraph dependencies and semantic conflicts between devices and applications. Traditional natural language processing (NLP) methods lack deep reasoning capabilities, while single large language models (LLMs) suffer from inherent reasoning biases and hallucination risks, failing to ensure strict legal traceability. To bridge this gap, we propose an evidence-traceable, mul tiagent collaborative compliance detection framework driven by heterogeneous LLMs. First, we construct a fine-grained evaluation paradigm based on the personal information protection law (PIPL) and employ an indicator-driven retrieval-augmented generation (RAG) mechanism to strictly anchor model reasoning to traceable policy fragments. Second, we design an evidence-quality-driven dynamic routing mechanism via a task complexity assessment model (TCAM), which adaptively dispatches multiple heterogeneous LLM agents based on quantitative difficulty and evidence scarcity, achieving an optimal balance between computational overhead and robustness. Furthermore, an iterative robust aggregation (IRA) and consistency-verification module is introduced to filter single-model biases and map identified risks to specific legal provisions. Extensive experiments demonstrate that the proposed framework achieves expert-level diagnostic precision with an average accuracy of 90.3%, a Macro-F1 of 89.3%, and a Cohen's Kappa of 0.85. Meanwhile, under the high-concurrence execution setting, the framework achieves approximately a 17-fold improvement in auditing efficiency compared with manual reviews, supporting high-throughput processing at 82 documents per hour. This end-to-end auditable system provides a highly scalable and interpretable regulatory technology solution for privacy governance in massive IoT ecosystems.

Item Type:	Article
Uncontrolled Keywords:	compliance detection,internet of things (iot),large language model (llm),multiagent system,personal information protection law (pipl),privacy policy,retrieval-augmented generation (rag),signal processing,information systems,hardware and architecture,computer science applications,computer networks and communications ,/dk/atira/pure/subjectarea/asjc/1700/1711
Faculty \ School:	Faculty of Science > School of Computing Sciences
Related URLs:	https://www.scopus.com/pages/publication...
Depositing User:	LivePure Connector
Date Deposited:	18 Sep 2026 09:22
Last Modified:	20 Sep 2026 05:26
URI:	https://ueaeprints.uea.ac.uk/id/eprint/104587
DOI:	10.1109/JIOT.2026.3697293

Downloads

Actions (login required)

View Item