Modelling Dependencies Between Classifiers in Mobile Masquerader Detection

The unauthorised use of mobile terminals may result in an abuse of sensitive information kept locally on the terminals or accessible over the network. Therefore, there is a need for security means capable of detecting the cases when the legitimate user of the terminal is substituted. The problem of user substitution detection is considered in the paper as a problem of classifying the behaviour of the person interacting with the terminal as originating from the user or someone else. Different aspects of behaviour are analysed by designated one-class classifiers whose classifications are subsequently combined. A modification of majority voting that takes into account some of the dependencies between individual classifiers is proposed as a scheme for combining one-class classifiers. It is hypothesised that by employing the proposed scheme, the classification accuracy may be improved as compared with the base majority voting scheme. The conducted experiments with synthetic data support this hypothesis.