Learning marked Markov modulated Poisson processes for online predictive analysis of attack scenarios

Runtime predictive analysis of quantitative models can support software reliability in various application scenarios. The spread of logging technologies promotes approaches where such models are learned from observed events. We consider a system visiting transient states of a hidden process until reaching a final state and producing observations with stochastic arrival times and types conditioned by visited states, and we abstract it as a marked Markov modulated Poisson Process~(MMMPP) with left-to right structure. We present an Expectation-Maximization (EM) algorithm that learns the MMMPP parameters from observation sequences acquired in repeated execution of the transient behavior, and we use the model at runtime to infer the current state of the process from actual observed events and to dynamically evaluate the remaining time to the final state. The approach is illustrated using synthetic datasets generated from a stochastic attack tree of the literature enriched with an observation model associating each state with an expected statistics of observation types and arrival times. Accuracy of prediction is evaluated under different variability of hidden states sojourn durations and of the observations arrival process, and compared against previous literature that mainly exploits either the timing or the types of observed events.