Adversarial Techniques in LPI Airborne Radar (ATILAR)

Vadum will develop Adversarial Techniques in LPI Airborne Radar (ATILAR) a novel waveform generation technique to counter machine learning-based detectors in modern cognitive EW receivers. ATILAR will integrate adversarial Machine Learning (ML) algorithms with cognitive radar RF coexistence techniques to enhance the ability of LPI radars to evade detection.  ATILAR will develop LPI waveforms to counter modern ML-based classifiers, which are susceptible to failure when confronted with adversarial noisy inputs. Additionally, RF spectral coexistence techniques using adaptive waveforms will be exploited to further enhance LPI performance by allowing the radar to embed its signals among other RF emitters in the same band. This combined approach will allow the Navy airborne LPI radars to regain their stealth advantage over adversary EW systems, so that they can detect targets of interest at tactically useful ranges while avoiding detection of their radar transmissions. The goal of this Phase I effort is to develop the algorithms for both adversarial LPI waveform creation and for adaptive use of the RF spectrum.  The ATILAR goal is to provide at least a 10X reduction in the performance of EW detector/classifiers, compared to a conventional LPI waveform, with less than a 10% degradation in the waveform’s radar performance.  The initial step of adversarial LPI waveform development is designed to fool the modern ML-based EW modulation classifiers, causing them to mis-classify the waveform.  The second step, adaptive use of the RF spectrum, is designed to embed the LPI waveform in an area of RF interference, making it harder to both detect and correctly classify the radar signal. The goal of this Phase I effort is to develop the algorithms for both adversarial LPI waveform creation and for adaptive use of the RF spectrum to help hide the radar signal.  The ATILAR goal is to provide at least a 10X reduction in the performance of EW detector/classifiers, compared to a conventional LPI waveform, with less than a 10% degradation in the waveform’s radar performance.  The initial step, adversarial LPI waveform creation, is designed to cause ML-based EW modulation classifiers to mis-classify the LPI waveform modulations.  The second step, making adaptive use of the RF environment characteristics, will embed the LPI waveform in a region of RF interference, enhancing the adversarial noise effects on the adversary EW system. The Vadum team has extensive experience in the application of machine learning techniques for Cognitive EW including unknown online radar threat characterization, tracking and threat behavior learning. Vadum will also utilize its transition and commercialization experience gained through successful execution and transition of several science and technology efforts to Programs of Record.