Toolbox for Verification of Autonomous Systems with Neural Components (TOOL-VAN)

As the Navy seeks to deploy neural networks for perception and control capabilities in autonomous systems, developers must ensure predictable and reliable network behavior. Researchers have produced a proliferation of methods for formally verifying the properties of neural networks, but applying these methods requires a prohibitively high level of technical expertise because they vary in their applicability to different architectures, properties, and domains. To address the need for user-friendly verification methods, Charles River Analytics designed and demonstrated the feasibility of a Toolbox for Verification of Autonomous Systems with Neural Components (TOOL-VAN), a general-purpose toolbox using automated verification method selection, specification-based retraining, cyber¢physical system integration, and an intuitive graphical user interface to allow users of different backgrounds to verify autonomous systems with neural components in safety, mission, and time-critical scenarios. In Phase I we showed TOOL-VANÆs ability to lower the barrier of entry to neural network verification, enabling the safe deployment of artificial intelligence systems in the Navy and throughout the Federal Government. Under this Phase II effort, we propose to expand the library of formal verification methods, further refine our intelligent user interface, develop a full-scope prototype, and demonstrate TOOL-VAN on integrated cyber¢physical systems with neural network components.