Guided Integration of Formal Verification in Assurance Cases

Abstract

Assurance cases are being increasingly acknowledged as a way to build trust in complex systems with autonomous capabilities. An assurance case is a comprehensive, defensible, and valid justification that a system will function as intended for a specific mission and operating environment. Formal verification is often reserved for the most critical components of such systems. However, formal verification tools are themselves often complex, and their usage is subject to many constraints and contextual dependencies. This can raise challenges both for performing the verification and reflecting the verification results appropriately in the assurance case, especially for non-expert users of the formal tool.

To address these challenges, we present a tool-supported methodology for integrating formal verification into an assurance case by capturing key verification method information in a rigorously constructed assurance case. In particular, we capture the tool specifications in terms of its inputs, outputs, and assurance constraints as assumptions over inputs and guarantees provided over its outputs. The tool specifications are parametrized over the inputs and outputs to both guide the intended application of the tool, and to check that the tool has been applied following the stated assumptions and that the guarantees hold. We define a generic tool assurance argument pattern that enables integration of the verification in the assurance case through custom refinement and automated instantiation for each tool use. We demonstrate our methodology on two formal verification tools and their applications to the verification of neural network properties for the aircraft domain.