The use of partially observable Markov decision processes to optimally implement moving target defense

For moving target defense (MTD) to shift advantage away from cyber attackers, we need techniques which render systems unpredictable but still manageable. We formulate a partially observable Markov decision process (POMDP) which facilitates optimized MTD capable of thwarting cyber attacks without excess overhead. This paper describes POMDP formulation including the use of an absorbing final state and attack penalty scaling factor to abstract defender-defined priorities into the model. An autonomous agent leverages the POMDP to select the optimal defense based on assessed cyber-attack phase. We offer an example formulation wherein attack suppression of greater than 99% and system availability of greater than 94% were maintained even as probability of detection of attack phase dropped to 74%.