Deep Learning Run Time Verifier (DL-RTV)

US Naval combat management systems are comprised of a variety of interconnected hardware architectures, firmware, and operating systems. As these systems are critical in nature, measures must be taken to identify any errors that may result in catastrophic system failure. However, a myriad of configuration combinations present due to hardware and software refresh cycles and update status result in a wide variety of failure modes that may not always be able to be fully tested prior to deployment. In response, we propose a Deep Learning Run Time Verifier (DL-RTV), which provides runtime verification that trains transformer models on vulnerable input patterns via an offline mode of operation and monitors combat management system behavior during online modes of operation. The offline mode of operation performs strategic fuzzing of virtualized combat management system devices to generate data for ingest by transformer models that train to identify those input patterns that lead to vulnerable states being reached in the target software. The online module ingests network traffic and system configuration parameters and applies the transformer model to identify triggers for Indications and Warnings of potentially threatening input vectors are identified through inference operations using the transformer model.