<![CDATA[Complex campus networks featuring server-based services and the growing Internet of Things (IoT) require near-real-time monitoring systems without incurring significant overhead. This study proposes a lightweight Artificial Intelligence-Internet of Things (AI-IoT)-based network monitoring prototype on an edge computing platform, utilizing an unsupervised autoencoder for anomaly detection. This prototype is implemented out-of-band on a Raspberry Pi 4 Model B device that serves as both a collection and inference node. The deep learning model on the TensorFlow Lite framework is compressed using TinyML for compatibility with small devices. The results use a dataset of 600,000 labeled flows that illustrate the trade-off in operational flexibility. At the P70 threshold, an F1-Score of 0.60 (precision 0.96, recall 0.43) is obtained, and in the P95 scenario, false positives can be completely eliminated. The edge infrastructure demonstrated excellent performance with an average batch processing latency of 74 ms and a throughput of over 300 flows/second with a constant Random Access Memory (RAM) usage of 2.8%.]]>
Copyrights © 2026
