Research Advances in Automated Red Teaming James Decraene, Fanchao Zeng, Malcolm Yoke Hean Low, Suiping Zhou, Wentong Cai Parallel and Distributed Computing Center School of Computer Engineering Nanyang Technological University, Singapore 639798 {jdecraene,fczeng,yhlow,asspzhou,aswtcai}@ntu.edu.sg Keywords: military agent-based modelling and simulation, automated red teaming, constraint handling, cloud comput- ing. Abstract We present, combine and apply novel research advances to Automated Red Teaming (ART). ART is an automated vul- nerability assessment tool which is employed to uncover the hard-to-predict and potentially critical elements of military operations. ART is principally based on the use of agent- based modelling/simulation, data farming and evolutionary computation. In this paper, we present two distinct computa- tional methods to address multiple issues of ART: constraint handling and computing budget. These novel techniques orig- inate from the research fields of evolutionary computation and cloud computing. These techniques are applied to a mil- itary toy model which was developed with the agent-based simulation platform MANA. We then discuss another poten- tial bottleneck of ART: many-objective optimization. The aim of this research is to optimize ART to best assist defense ex- perts in operational analysis and, ultimately, in critical deci- sion making. 1. INTRODUCTION To assist the military decision making process, Red Team- ing (RT) [22, 28] was proposed as a vulnerability assessment tool which enables one to uncover the hard-to-predict and po- tentially disruptive elements of a military operation. Using this method, defense analysts may subsequently identify and resolve the weaknesses of tactical plans/defense systems. In typical (human-based) RT simulations, a defensive blue team is subjected to repeated attacks, where multiple scenarios may be examined, from a belligerent red team. RT has proved to be a valuable method to improve the robustness of operational tactics such as protecting key facilities (e.g., nuclear plants, military installations, etc.) [12]. Nevertheless RT is highly time-consuming where only a limited range of scenarios may be investigated due to practi- cal constraints. Automated Red Teaming (ART) [26, 6] (or “objective-based data farming” [7]) was proposed to over- come this limitation by addressing RT in an automated man- ner using Evolutionary Agent-Based Simulations (EABS) [14]. EABSs are computational methods which can model the intricate and non-linear dynamics of warfare. EABSs utilize Evolutionary Computation (EC) techniques [10] to evolve simulation models to exhibit pre-specified/desirable output behaviors. Moreover, these EC techniques are commonly de- vised to solve multi-objective optimization problems as mili- tary operations are characterized with such multi-dimensional constraints which often conflict with each other. In ART, the parameter values (e.g., troop cluster- ing/cohesion, response to injured teammates, aggressiveness, stealthiness, etc.) defining the behavior or personality of the red team are evolved to optimize its efficiency (e.g., maximize damage to target facilities) against the blue team. Examples of ART systems which have been applied to military decision making include: ISAAC/EINSTein [14], WISDOM [28] and NALEX [24]. Although ART has successfully been applied to a variety of military studies, we argue that significant issues still exist when using this tool. These issues are identified as follows: 1. Constraint handling: As mentioned earlier, specific pa- rameter values of the red team are subjected to evolution in ART. Nevertheless no trade-off in cost has been intro- duced, as a result the evolutionary process may vary the parameter values regardless of their financial or practical cost. For example in [20], the optimal red team configu- ration was found with a value of 97% for “stealthiness”, this imposes hard practical constraints to real-life oper- ations which may not always be easily achieved. More- over, the evolutionary process may result in simulation models which are not plausible or more critically, valid. For example, we may consider the evolution of agents’ spatial coordinnates in a three dimensionnal space, the ART process may result in positioning agents at invalid locations (e.g., within a wall or in the sky). This may dramatically limit the potential of ART in such circum- stances. 2. Computing budget: ART (and more generally data farm- ing) experiments typically require high performance computing facilities which availability and capabilities may not satisfy the user’s time constraints and experi- mental requirements. The ART user may thus be con- fronted with a “computing budget” issue which may