Date of Award
Doctor of Information Technology
School of Computer and Security Science
Computing, Health and Science
Associate Professor Chiou-Peng Lam
Associate Professor Philip Hingston
Dr Martin Masek
Red teaming (RT) is a process that assists an organization in finding vulnerabilities in a system whereby the organization itself takes on the role of an “attacker” to test the system. It is used in various domains including military operations. Traditionally, it is a manual process with some obvious weaknesses: it is expensive, time-consuming, and limited from the perspective of humans “thinking inside the box”. Automated RT is an approach that has the potential to overcome these weaknesses. In this approach both the red team (enemy forces) and blue team (friendly forces) are modelled as intelligent agents in a multi-agent system and the idea is to run many computer simulations, pitting the plan of the red team against the plan of blue team.
This research project investigated techniques that can support automated red teaming by conducting a systematic study involving a genetic algorithm (GA), a basic coevolutionary algorithm and three variants of the coevolutionary algorithm. An initial pilot study involving the GA showed some limitations, as GAs only support the optimization of a single population at a time against a fixed strategy. However, in red teaming it is not sufficient to consider just one, or even a few, opponent‟s strategies as, in reality, each team needs to adjust their strategy to account for different strategies that competing teams may utilize at different points. Coevolutionary algorithms (CEAs) were identified as suitable algorithms which were capable of optimizing two teams simultaneously for red teaming. The subsequent investigation of CEAs examined their performance in addressing the characteristics of red teaming problems, such as intransitivity relationships and multimodality, before employing them to optimize two red teaming scenarios. A number of measures were used to evaluate the performance of CEAs and in terms of multimodality, this study introduced a novel n-peak problem and a new performance measure based on the Circular Earth Movers‟ Distance.
Results from the investigations involving an intransitive number problem, multimodal problem and two red teaming scenarios showed that in terms of the performance measures used, there is not a single algorithm that consistently outperforms the others across the four test problems. Applications of CEAs on the red teaming scenarios showed that all four variants produced interesting evolved strategies at the end of the optimization process, as well as providing evidence of the potential of CEAs in their future application in red teaming.
The developed techniques can potentially be used for red teaming in military operations or analysis for protection of critical infrastructure. The benefits include the modelling of more realistic interactions between the teams, the ability to anticipate and to counteract potentially new types of attacks as well as providing a cost effective solution.
Ranjeet, T. (2012). Coevolutionary algorithms for the optimization of strategies for red teaming applications. Retrieved from http://ro.ecu.edu.au/theses/558