Real-Time Detection of Spot Jamming Attacks in mmWave Radar Systems Using a Lightweight CNN

Abstract

Millimeter-wave (mmWave) radars are integral to advanced driver assistance systems for object detection and tracking. However, these radars are vulnerable to interference from other mmWave radars in the vicinity, potentially leading to false detections and tracking errors. This letter focuses on identifying which frames of ego radar data are affected by spurious signals from a spot jamming attack (a scenario where one radar intentionally interferes with another with the same specifications). We conducted experiments using two AWR1843 radars, with one acting as the jammer, and observed only a few frames of data were falling under a spot jamming attack. We transformed the in-phase and quadrature-phase (I-Q) data from the ego radar into range-angle heatmap images using 2-D fast Fourier transform (2D-FFT). On 2D-FFT images, a lightweight convolution neural network (CNN) classifier with a model size of 5MB is proposed to distinguish between jammed and nonjammed frames. The classifier exhibits a 95.4% accuracy in ten-fold cross-validation, outperforming pretrained models, such as DenseNet, EfficientNet, InceptionNet, MobileNet, NASNet, ResNet, VGGNet, ConvNeXt, and Xception. Moreover, the CNN was successfully deployed on edge devices, Raspberry Pi, and other processors, observing the execution of CNN in just 15.8 milliseconds per frame. This work demonstrates the potential for real-time detection of spot jamming attacks, with applications in electronic counter-countermeasures, source localization, machine learning (ML)-aided passive radar systems, and cognitive radar development.