Adaptive LPD Radar Waveform Design with Generative Deep Learning

TL;DR

The paper tackles the challenge of designing low probability of detection (LPD) radar waveforms that blend into ambient RF background without sacrificing sensing performance. It introduces a conditional Wasserstein GAN with gradient penalty (cWGAN-GP) to learn background-matching waveform distributions, conditioned on instantaneous RF background $y$, and couples this with an ambiguity-function based loss that favors a thumbtack-like $\hat{A}(\tau,F_D)$. The ambiguity loss comprises a mainlobe term and a sidelobe term (weighted for zero Doppler) so that generated waveforms maintain useful range/velocity resolution while remaining hard to detect; the total objective is $L_{total}=L_W+\eta L_{ambig}$. Experiments on toy LFM chirps and the RadioML/ SIDLE datasets show the approach can reduce detectability by up to 90% relative to traditional LPD waveforms, while achieving favorable sensing metrics; the framework also offers a tunable trade-off between detectability and sensing by adjusting $\eta$ and conditioning on the RF background.

Abstract

We propose a learning-based method for adaptively generating low probability of detection (LPD) radar waveforms that blend into their operating environment. Our waveforms are designed to follow a distribution that is indistinguishable from the ambient radio frequency (RF) background -- while still being effective at ranging and sensing. To do so, we use an unsupervised, adversarial learning framework; our generator network produces waveforms designed to confuse a critic network, which is optimized to differentiate generated waveforms from the background. To ensure our generated waveforms are still effective for sensing, we introduce and minimize an ambiguity function-based loss on the generated waveforms. We evaluate the performance of our method by comparing the single-pulse detectability of our generated waveforms with traditional LPD waveforms using a separately trained detection neural network. We find that our method can generate LPD waveforms that reduce detectability by up to 90% while simultaneously offering improved ambiguity function (sensing) characteristics. Our framework also provides a mechanism to trade-off detectability and sensing performance.

Figures (14)

• Figure 1: LPD Radar Waveform Generation Framework. The generative model optimizes range/velocity resolution while minimizing detectability.
• Figure 2: Thumbtack Ambiguity Function. Example of an optimal "thumbtack" ambiguity function with low delay (range) and Doppler (velocity) ambiguities.
• Figure 3: Network Architectures. Overview diagrams of the generator network (left) and critic network (right) with 1024 signal length and 512 latent vector length. Output shapes are shown below layers with batch size $B$.
• Figure 4: Adversarial Training Framework. Training process for our cWGAN-GP, where our generator learns to produce waveforms with desirable ambiguity functions that mimic the RF background. Training is initially performed without the ambiguity loss, and alternates between training the critic and the generator. Once converged, the generator is fine-tuned for a small number of epochs with both the cWGAN-GP loss and ambiguity loss.
• Figure 5: Evaluation Framework. Training process for the detector neural network used to evaluate LPD performance. We add two additional classes---generated and empty---to the baseline waveforms to create the combined dataset. We add randomly time-shifted RF background signals to these examples using waveforms from RadioML 2018.01A that have not been seen by our generator network. We evaluate detectability with an FFT Accumulation Method (FAM)-based cyclostationary detector and a neural network-based detector.