MSVIB-Net: Taming High-Dimensional GNSS Signals for Real-Time Spoofing Detection on AIoT Devices

Jing Ji, Quanguo Ye, Wei Chen, Changzhen Li, Luyao Du, Zhonghui Pei, Hongyang Lu, Jiantong Zhang

Peer Reviewed

Abstract:	Global Navigation Satellite System (GNSS) spoofing presents a growing threat to the integrity of positioning, navigation, and timing (PNT) services, with emerging Artificial Intelligence of Things (AIoT) devices being particularly vulnerable due to their stringent resource constraints and reliance on edge computing. Traditional spoofing detection algorithms often face a critical trade-off: complex, highdimensional signal processing achieves high accuracy at the cost of computational latency, while lightweight methods sacrifice sensitivity. To break this trade-off, we propose MSVIB-Net, a novel lightweight detector architected for real-time performance on AIoT devices. MSVIB-Net synergistically integrates a Convolutional Neural Network (CNN) for extracting spatial correlations in the signal spectrum, a Long Short-Term Memory (LSTM) network for modeling temporal dependencies, and a core Minimum Sufficient Variational Information Bottleneck (MSVIB) module. The MSVIB module is pivotal, it operates as an intelligent, learnable compressor that distills high-dimensional GNSS signal data into a minimal sufficient statistic, ruthlessly eliminating redundant and noisy information while preserving only the features most critical for classification. This process of information compression not only enhances robustness against subtle and evolving spoofing strategies but also drastically reduces the computational footprint of the subsequent network layers. Evaluated extensively on the open-source OAKBAT dataset, MSVIB-Net demonstrates state-of-the-art performance, achieving over 99% detection accuracy. More significantly for deployment, it yields an 82.58% reduction in training time compared to a standard CNN-LSTM baseline. In inference, it achieves a 60% reduction in latency and a 26.11% decrease in memory usage over a competitive Generative Adversarial Network (GAN)-based detector. The MSVIB module itself is shown to boost classification performance on mixed authentic and spoofed signals by at least 15%. This work concludes that deliberate, learned feature compression via the Information Bottleneck principle is the key to enabling robust, real-time GNSS security on the edge. MSVIB-Net establishes a new paradigm for resource-frugal AI-driven signal authentication, directly advancing the practical deployment of secure PNT in next-generation AIoT ecosystems.
Published in:	Proceedings of the 2026 International Technical Meeting of The Institute of Navigation January 26 - 29, 2026 Hyatt Regency Orange County Anaheim, California
Pages:	22 - 38
Cite this article:	Ji, Jing, Ye, Quanguo, Chen, Wei, Li, Changzhen, Du, Luyao, Pei, Zhonghui, Lu, Hongyang, Zhang, Jiantong, "MSVIB-Net: Taming High-Dimensional GNSS Signals for Real-Time Spoofing Detection on AIoT Devices," Proceedings of the 2026 International Technical Meeting of The Institute of Navigation, Anaheim, California, January 2026, pp. 22-38. https://doi.org/10.33012/2026.20515
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In