GRAF RESEARCH CORPORATION — Department of Defense SBIR Phase I: AF221-0022

Description

While there is extensive literature defining approaches and applications of artificial intelligence, the field of AI explainability is relatively new. Understanding the reasoning behind a logical result can provide a level of validity and reliability that may not be readily apparent.  In recent years, radio frequency communications have used machine learning. Research has moved from classical, iterative signal classification algorithms to applications of Deep Neural Networks. RFML implementations will continue to improve in terms of classification performance and real-time classification capability. AI algorithms, however, need to be explainable to allow for sufficient levels of Test and Evaluation for a system to be declared “operational” and be deployed with the warfighter in radar, electronic warfare, ELINT, and SIGINT operations. This effort will commence with the analysis of explainability methodologies and novel metrics for AI analysis applied to the RF domain.  The envisioned process will establish feasibility and plan for the development of a means to explain and measure AI  for RF data.  The research program focuses on producing a unique set of explainability AI methodologies and metrics, a quantification of the validity and reliability of the RFML explainability method, and an exploration of the needs for near real-time implementation to be applicable to fielded RFML systems.