Rigorous Machine Learning for Secure and Autonomous Cyber Physical Systems

摘要

Machine learning (ML) based secure and autonomous cyber physical systems are often not reliable and interpretable mainly because the employed ML techniques suffer from false alarms that may result in physical and financial loss. We assert that reliability and interpret-ability of the ML methods depends on underlying statistical models that infer results. Therefore, we introduce a rigorous method for the model selection. Current selection methods choose a model using statistical criteria (e.g., AIC, BIC). These criteria may lead to selection of an inappropriate model (e.g. over/under-fitting) because they only consider relative-quality (statistical) of the model without considering absolute-quality (formal) of the model based on the model/data specification. To this end, we argue the suitability of recently developed-decidability procedures/solvers. Such solvers infer if a selected model can(not) classify a given data and produce a formal proof that can be used to assure reliability and security of modelled system. We demonstrate feasibility of the method through a simple example of an autonomous insulin pump.