Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

MSVIB-Net: A Physically-Informed Lightweight AI Framework for Real-Time GNSS Spoofing Detection
Jing Ji, VCSCIT, Wuhan Xingtu Xinke Electronics Co., Ltd./ School of Information Engineering, Wuhan University of Technology; Quanguo Ye, Wei Chen, School of Automation, Wuhan University of Technology; Changzhen Li, School of Information Engineering, Wuhan University of Technology; Luyao Du, School of Automation, Wuhan University of Technology; Zhonghui Pei, School of Computer Science and Engineering, Wuhan Institute of Technology; Hongyang Lu, China Transport Telecommunications & Information Center (CTTIC) / School of Information Engineering, Wuhan University of Technology; Jiantong Zhang, China Transport Telecommunications & Information Center (CTTIC)
Location: Windsong 7-8

Global Navigation Satellite System (GNSS) spoofing attacks pose a critical threat to infrastructure and autonomous systems. While deep learning detection methods show promise, their deployment is constrained by excessive complexity, lack of physical interpretability, and insufficient recall. This paper introduces MSVIB-Net, a physically-informed lightweight AI framework that integrates spatiotemporal deep learning with information-theoretic optimization. The core Minimum Sufficient Variational Information Bottleneck (MSVIB) model distills high-dimensional GNSS signals into a minimal yet maximally discriminative latent representation, guided by physical principles validated on the OAKBAT dataset across five spoofing scenarios. The hybrid CNN-LSTM architecture extracts spatial and temporal features, compressed by MSVIB to very few dimensions. Results demonstrate exceptional performance with 99.19% accuracy, 0.9993 recall (0.07% miss rate), a 30.1% improvement over baselines, with extreme lightweight efficiency—27.8 KB memory footprint, 22.04 ms inference latency, and 0.4 s detection delay—enabling direct deployment on resource-constrained edge platforms. The framework maintains robustness under noise and adversarial attacks. Crucially, its physical interpretability—features directly traceable to specific attack mechanisms—enhances explainability and certifiability, enabling integration with standards like OSNMA. This work unifies information theory, physical modeling, and lightweight ML to establish a new paradigm for certifiable, deployable GNSS spoofing detection for resilient PNT in the autonomous systems era.