Description
Anomaly detection is a big part of many areas of our daily lives, such as financial
services, network monitoring, and medical diagnosis. Most of the application fields are
either financially, socially, or security-relevant domains, where precise anomaly/outlier
detection methods are important for the successful operation of these services. Unsurprisingly,
the field of anomaly detection on classical data is highly researched, bringing
up many statistical and machine-learning approaches. With the recent appearance of
quantum machine learning offering interesting possible advances in computational
power and latent space complexity, the next step is to look into possible methods for
anomaly detection using this approach. We tackle the anomaly detection problem in
time series data by building upon existing methods, identifying their weaknesses, and
proposing new model architectures to improve on them. This thesis introduces two
model formulations, the window model and the multi-time-step model, which are more
sensitive to the trained data by giving them better temporal insight into the data. Most
time-dependent data is influenced by its past, creating interesting anomaly scenarios
that are difficult to detect.
Our findings emphasize the importance of the temporal insights the model can gain
through reformulating the model’s architectures. The evaluations show a significant
improvement compared to the reference method on all tested scenarios, particularly
on collective anomalies spanning multiple time steps, where the improved temporal
knowledge leads to higher anomaly detection accuracies.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching