Reliable AI Explanations for Testing and Evaluation of RF Systems (RAETERS)

The application of artificial intelligence (AI) and machine learning (ML) techniques, such as deep learning, to radio frequency (RF) systems, opens enormous new opportunities to improve communications, sensors, intelligence gathering, and electronic warfare. One major challenge is to figure out how to effectively test and evaluate (T&E) such systems. With sufficient knowledge and expertise, it is generally possible to predict how conventional RF systems will behave given a certain set of conditions and to establish verification, validation, and accreditation (VV&A) policies based on that knowledge. For deep learning-based systems, it is extremely difficult for a human to predict what the output will be, making it extremely difficult to develop appropriate VV&A policies. Charles River Analytics proposes to design and prototype Reliable AI Explanations for Testing and Evaluation of Radiofrequency Systems (RAETERS), a test and evaluation capability to establish the performance and reliability of machine learning-based radiofrequency systems using a multi-stage adversarial model inversion approach to explain the behavior of the system under test and establish performance metrics with operationally relevant simulations. RAETERS will use adversarial model inversion to generate explanations and calculate performance and reliability metrics for deep learning-based RF systems to establish that such systems operate as intended.