Lightweight physical-layer authentication for IoT devices access against jamming attacks

Abstract

Radio frequency fingerprint identification (RFFI) is a reliable non-cryptographic physical layer security technique that leverages transmitter hardware features for device identification. However, Internet of Things (IoT) environments often contain various interference sources, which can cause physical layer access authentication failures. Furthermore, RFFI systems typically require substantial computational resources. To reduce computing resources, we propose a lightweight and scalable IoT wireless signal identity detection system against jamming attacks. We propose SE-MobileNet, a lightweight convolutional neural network that enhances MobileNetV2 by replacing its repetitive bottleneck blocks with squeeze-and-excitation (SE) modules, thereby augmenting channel-wise feature recalibration and improving representational capacity. Specifically, the received signals are firstly transformed into channel independent spectrograms through multiple processing steps, and fed into SE-MobileNet to train the radio frequency fingerprint (RFF) feature extractor. Then, the RFF feature extractor fetches the RFF features of the devices and uses k-nearest neighbor (KNN) to complete abnormal device detection and legitimate device classification. Experimental results show that the system model occupies only 2.41MB, representing a 99% reduction compared to existing benchmarks. We can achieve the highest area under curve (AUC) value of 0.971 for abnormal device detection and 97.6% accuracy rate for legitimate device classification. Even under co-channel interference attacks, SE-MobileNet maintains detection accuracy above 0.70 in weak interference environments and achieves or exceeds 0.95 at moderate-to-high SIR levels. This indicates that the model can maintain high accuracy despite significant model compression.