A Tiny Machine Learning for Real-Time Anomaly Detection of Self-Media Public Opinion in Edge-Cloud-Cooperation Campus Networks

Abstract

Real-time anomaly detection in self-media public opinion requires lightweight solutions to address the latency and multimodal complexity challenges of campus network ecosystems. This article proposes a tiny machine learning framework for edge-cloud-cooperation campus networks, enabling efficient detection of opinion anomalies through distributed computation. The architecture combines edge-native micro-model compression with cloud-assisted federated verification, achieving three key innovations: (1) On-device micro-graph neural networks (GNNs) deployed at edge nodes for low-latency pattern recognition in terahertz multimedia streams; (2) a dual-phase anomaly engine leveraging contrastive semantic alignment and adaptive influence analysis to capture cross-modal inconsistencies; (3) dynamic knowledge distillation that reduces model footprints to 8 MB while preserving 91% precision and 87% recall on a 120,000 post dataset from 15 universities. Experimental results demonstrate 120 ms average inference latency with 68% lower computation overhead than centralized baselines, accelerating emergency response by 3.25× through edge-cloud task partitioning. The framework maintains 74% energy efficiency in continuous operation, proving the viability of tiny machine learning paradigms for intelligent campus governance without relying on next-generation communication standards.