Anomaly Detection on Tiny Devices

• During this project you will work at the intersection of three areas:
  Resource-constrained embedded devices for Internet of Things (IoT);
• Tiny machine learning (TinyML);
• Anomaly detection.

The project aims to work on the task of anomaly detection, that is the identification of rare events or observations which differ significantly from standard patterns present in data streams. It is common to build anomaly detection models using data-driven techniques and dedicated machine learning methods. However, this approach assumes access to powerful machines, which is not granted: the special scenario to be investigated in this project, is the setting where the anomaly detection models need to be deployed on embedded devices with limited storage and compute (i.e., resource-constrained), a common requirement within the IoT context. Design and deployment of such models is within the realm of
TinyML, which explicitly aims at studying machine learning methods that provide strong performance under severe resource constraints. To evaluate the efficacy of TinyML methods, researchers constructed dedicated benchmarks. Within this project, you will evaluate your solution on the anomaly detection dataset that is a part of MLPerf Tiny Benchmark.

Key objectives

•  Explore and implement TinyML methods for anomaly detection;
• Integrate the developed methods with MLPerf Tiny Benchmark’s dataset for anomaly detection;
• Evaluate the performance and scalability of the methods, assessing the trade-off between the quality of anomality detection and storage and compute costs.

Supervisors

Dr. Denis Kleyko: denis.kleyko@oru.se
Dr. Alberto Giaretta: alberto.giaretta@oru.se