Possession of modern smartphones is becoming increasingly ubiquitous, and with this rise in usage comes a rise in the amount of sensitive data being stored on them. Despite this, the high-frequency, low-duration nature of the average smartphone session makes passwords or PIN-locks even less usable than in the desktop context. To combat these issues, implicit authentication (IA) schemes can be developed and deployed to smartphones. IA schemes continuously authenticate users by profiling their behaviour using the variety of sensors prevalent on the phones, such as touchscreens and accelerometers. When a non-owner acquires the device and attempts to access sensitive data on it, the IA scheme recognizes the difference in behaviour and automatically ejects the attacker from the system. In particularly sensitive contexts, IA schemes can also be deployed as a secondary defence mechanism on top of explicit authentication, providing layered security in the event of, for example, a shoulder-surfing attack compromising the device's PIN or an operating system vulnerability allowing its bypass. In this work, we evaluate existing proposals for IA schemes using different behavioural feature sets, and evaluate them against real-world data to show when they are (and are not) useful. We have implemented them in an easily extensible open source framework for the Android operating system called Itus, which allows other researchers to iteratively improve on the existing mechanisms for performing IA. Itus performs IA at the app level, which we have shown allows app developers to selectively protect sensitive data while decreasing the impact on battery life and device performance, and at the same time obtaining better detection accuracy for the IA scheme being invoked.

Security and usability issues with pass-locks on mobile devices have prompted researchers to develop implicit authentication (IA) schemes, which continuously and transparently authenticate users using behavioural biometrics. Contemporary IA schemes proposed by the research community are challenging to deploy, and there is a need for a framework that supports: different behavioural classifiers, given that different apps have different requirements; app developers using IA without becoming domain experts; and real-time classification on resource-constrained mobile devices. We present Itus, an IA framework for Android that allows the research community to improve IA schemes incrementally, while allowing app developers to adopt these improvements at their own pace.

A framework for coevolving genetic programming teams with Pareto archiving is benchmarked under two representative tasks for non-stationary streaming environments. The specific interest lies in determining the relative contribution of diversity and aging heuristics to the maintenance of the Pareto archive. Pareto archiving, in turn, is responsible for targeting data (and therefore champion individuals) as appropriate for retention beyond the limiting scope of the sliding window interface to the data stream. Fitness sharing alone is considered most effective under a non-stationary stream characterized by continuous (incremental) changes. Fitness sharing with an aging heuristic acts as the preferred heuristic when the stream is characterized by non-stationary stepwise changes.

Classification as applied to streaming data implies that only a small number of new training instances appear at each generation and are never explicitly reintroduced by the stream. Pareto competitive coevolution provides a potential framework for archiving useful training instances between generations under an archive of finite size. Such a coevolutionary framework is defined for the online evolution of classifiers under genetic programming. Benchmarking is performed under multi-class data sets with class imbalance and training partitions with between 1,000's to 100,000's of instances. The impact of enforcing different constraints for accessing the stream are investigated. The role of online adaptation is explicitly documented and tests made on the relative impact of label error on the quality of streaming classifier results.