Main article

Yifan Chen
School of Computer Science, Nanjing University of Information Science and Technology, China 210044
Ibrahim Al-Najjar*
Department of Information Systems, University of Sharjah, Sharjah, United Arab Emirates 27272
ibrahim.alnajjar@sharjah.ac.ae

Abstract

This article develops a technical framework for self-supervised behavioral risk monitoring of large language models deployed in edge intelligence environments. Inspired by recent work on deception detection for Edge-of-Things systems, the paper broadens the problem from narrow deception labels to a wider class of trust failures that include concealed unsafe plans, hallucinated justifications, unstable tool use, retrieval mismatch, and context-dependent policy evasion. The core argument is that trustworthy edge deployment requires more than compressed inference; it requires a native monitoring layer that can transform local reasoning traces, uncertainty cues, and tool telemetry into actionable risk signals without relying continuously on cloud-based judge models. The paper synthesizes research on edge computing, language-model alignment, quantization, chain-of-thought reasoning, retrieval augmentation, and intrusion detection, and uses that synthesis to propose a layered architecture composed of a primary reasoning model, a lightweight monitor head, a telemetry collector, and a policy guard. It further outlines implementation pathways, evaluation metrics, and governance implications for privacy-sensitive and latency-critical domains. The contribution lies in showing how self-supervised trust monitoring can become a practical design pattern for edge AI when computational efficiency, behavioral reliability, and organizational accountability are engineered together.

Article details

How to Cite

Chen, Y. ., & Al-Najjar, I. . (2025). Self-Supervised Behavioral Risk Monitoring for Large Language Models in Edge Intelligence Environments. Journal of AI Analytics and Applications, 3(3), 1-18. https://doi.org/10.63646/jaiaa.2025.030301