RADGAN: Applying Adversarial Machine Learning to Track-Before-Detect Radar

In this paper, the authors explore employing Generative Adversarial Networks (GANs) toward radar target detection. Current machine learning approaches focus on Convolutional Neural Networks (CNNs) for real-time detection. Most GANs do not require labeled data, and this paper investigates the use of unsupervised learning methods for detecting radar targets. The approach combines image processing techniques with a track-before-detect inspired image data. The networks can detect Swirling 0 targets at single-dwell output signal-to-noise ratios of 3 dB at a false alarm rate of 10
^-2
. Consequently, GANs may provide an efficient pre-screening approach to reduce the computational requirements of traditional track-before-detect approaches.