Multivariate anomaly detection in low compute settings

The collection of data feeds from networked and other devices is now ubiquitous. Such data
streams provide a unique opportunity to monitor each device's operational performance in
a non-intrusive way, with a view to providing early identification of potentially anomalous
behaviour. In response to this, there has been a flurry of activity in the development of
statistically-founded, computationally efficient anomaly detection methods during the last 4
years. Whilst these methods perform well in many settings, it is clear that a new approach is
needed in low compute settings where constraints in both computational power and energy
usage may arise. This is particularly true when multiple features are being monitored over
time. This project will seek to develop statistical theory and methods for energy-efficient anomaly
detection in this multivariate setting, that take these constraints into account and provide
guarantees on the features which can be detected.