Semi-Supervised Anomaly Detection Using One-Class RBF Networks

Anomaly detection can be defined as ”the problem of finding patterns in data that do not conform to expected behavior” (Chandola et al., 2009) and has been used in real world applications such as fraud detection (Ghosh and Reilly, 1994), monitoring healthcare (¿Sabi´c et al.), intrusion detection (Ghosh et al., 1998), and crowd motion analysis (Zhou et al., 2016). The main challenge in anomaly detection is the imbalanced nature of the training data, since obtaining data belonging to the anomalous class is usually expensive, if not impossible, while the data belonging to the normal class is much more accessible. Hence, since the training data is almost always dominated by the data instances of the normal class, anomaly detection is sometimes framed as the one-class classification problem. This thesis investigates the possibility of modifying the Radial Basis Function (RBF) networks into effective one-class RBF networks for the task of anomaly detection. RBF networks are selected as the foundation for this thesis due to their attractive properties, which include but are not limited to: being fast and efficient, being interpretable (Jin and Sendhoff, 2003) and being adaptable to dynamic changes (Liu et al., 2020; Han et al., 2011; Rapaka et al., 2003) in the distribution of the data with respect to time. A shallow one-class RBF network is proposed, and then by eliminating certain restrictions, a stronger and more flexible version of these networks is also developed. Finally, a strategy for deepening one-class RBF networks is proposed so that they can work with raw high-dimensional datasets, while benefiting from the automatic feature extraction capability of deep learning. Through evaluation experiments it is shown that these newly proposed approaches lead to anomaly detection models that match or exceed the performance of current state of the art approaches while maintaining the advantages of RBF networks