Discovering and Utilising Expert Knowledge from Security Event Logs

Security assessment and configuration is a methodology of protecting computer systems from malicious entities. It is a continuous process and heavily dependent on human experts, which are widely attributed to being in short supply. This can result in a system being left insecure because of the lack of easily accessible experience and specialist resources. While performing security tasks, human experts often revert to a system's event logs to determine status of security, such as failures, configuration modifications, system operations etc. However, finding and exploiting knowledge from event logs is a challenging and time-consuming task for non-experts. Hence, there is a strong need to provide mechanisms to make the process easier for security experts, as well as providing tools for those with significantly less security expertise. Doing so automatically allows for persistent and methodical testing without an excessive amount of manual time and effort, and makes computer security more accessible to on-experts. In this thesis, we present a novel technique to process security event logs of a system that have been evaluated and configured by a security expert, extract key domain knowledge indicative of human decision making, and automatically apply acquired knowledge to previously unseen systems by non-experts to recommend security improvements.

The proposed solution utilises association and causal rule mining techniques to automatically discover relationships in the event log entries. The relationships are in the form of cause and effect rules that define security-related patterns. These rules and other relevant information are encoded into a PDDL-based domain action model. The domain model and problem instance generated from any vulnerable system can then be used to produce a plan-of-action by employing a state-of-the-art automated planning algorithm. The plan can be exploited by non-professionals to identify the security issues and make improvements. Empirical analysis is subsequently performed on 21 live, real world event log datasets, where the acquired domain model and identified plans are closely examined. The solution's accuracy lies between 73% - 92% and gained a significant performance boost as compared to the manual approach of identifying event relationships.

The research presented in this thesis is an automation of extracting knowledge from event data steams. The previous research and current industry practices suggest that this knowledge elicitation is performed by human experts. As evident from the empirical analysis, we present a promising line of work that has the capacity to be utilised in commercial settings. This would reduce (or even eliminate) the dire and immediate need for human resources along with contributing towards financial savings.