Anomaly Detection on the Edge: Comparison of Reconstruction and Classification Based Approaches

Ł. Grzymkowski, T. Cejrowski (Arrow Electronics, Poland), T. Stefański (Gdansk University of Technology, Poland)

In this work we discuss and evaluate different approaches to solving anomaly detection task when the target platform is a tiny microcontroller. We investigate modeling techniques and propose a comprehensive set of measurements to analyze performance, compute and memory requirements, and power efficiency. We run experiments to collect these measurements on platforms used in TinyML systems including Cortex-M7, Cortex-M55 and Ethos-U55 running TensorFlow Lite for Microcontrollers. The measurements are collected for an autoencoder in reconstruction-based anomaly detection and a MobileNetV2-like model trained for classification. We show which approach is more suitable depending on the system requirements and constraints. This work underscores the need for a holistic approach in selecting modeling and deployment strategies, providing empirical evidence to guide the development of efficient on-device anomaly detection systems.