In this talk, we’ll discuss a new approach to designing radar waveforms that are difficult to detect. Our method leverages a learning-based framework to produce waveforms that blend into the ambient radio frequency (RF) environment, thereby reducing their probability of detection. These waveforms are simultaneously designed to maintain their effectiveness in ranging and sensing tasks. We utilize an unsupervised adversarial learning model consisting of a generator network that creates the waveforms and a critic network trained to differentiate these generated waveforms from the natural RF background. To ensure that our waveforms remain functional for sensing, we implement an optimization objective based on the ambiguity function.

Our evaluations show that our approach can significantly lower the single-pulse detectability of these low probability of detection (LPD) waveforms by up to 90% compared to traditional methods while maintaining or improving their sensing capabilities. Moreover, our method allows for a tunable trade-off between detectability and sensing effectiveness, offering a flexible solution for adapting to different operational requirements.