One-class learning with an Autoencoder Based Self Organizing Map

Abstract

Cyber-security defense techniques have begun to transcend from one-size fits all approach to personalized solutions. These techniques factor in user autonomy by monitoring the temporal and spatial changes in the user's behaviour. With time, such a system is known to develop a comfort to the user's interaction with the device. The motivation of this research is to enable a device to differentiate its owner from another user, by discovering behavioural patterns in the contextual and physiological information of smartphone usage. Naturally, this poses a one-class learning constraint. The proposed framework quantifies namely, (i) the dissimilarity in behaviours among any two users (ii) the exclusivity of each user's behaviour (inclass) from the world (outclass). The crucial aspect of this framework is to construct a representation of the most important properties of each user. To this end, the utility of a feed forward multilayer perceptron (MLP) in identifying an encoding that rebuilds the input data with least loss is examined. The claim is that such an encoding step poses improved data representations prior to clustering, a data description technique. However, both the encoding and clustering steps respect the one-class learning restriction i.e. relative to a single user. The evaluations on publicly available smartphone datasets, show that the resulting (user specific) behavioural models are capable of uniquely identifying each user. In particular, encoded contextual information are better anchors to behaviour modelling in comparison with encoded physiological information of smartphone data.