Bayesian network conflict detection for normative monitoring of black-box systems

Abstract

Bayesian networks are interpretable probabilistic models that can be constructed from both data and domain knowledge. They are applied in various domains and for different tasks, including that of anomaly detection, for which an easy to compute measure of data conflict exists. In this paper we consider the use of Bayesian networks to monitor input-output pairs of a black-box AI system, to establish whether the output is acceptable in the current context in which the AI system operates. A Bayesian network-based prescriptive, or normative, model is assumed that includes context variables relevant for deciding what is or is not acceptable. We analyse and adjust the conflict measure to make it applicable to our new type of monitoring setting.