Reinforcement Learning for GNSS Spoofing Detection: A Multi-Class DQN Approach with TEXBAT

Abstract

Global Navigation Satellite Systems (GNSS) are critical for positioning and timing, but the weak and unencrypted nature of civil GNSS signals makes them highly vulnerable to spoof-ing attacks. Existing detection methods, often based on handcrafted metrics or supervised learning, lack robustness across diverse scenarios. In this paper, we formulate GNSS spoofing detection as a multi-class reinforcement learning problem and propose a Deep Q-Network (DQN) trained on GNSS tracking features from the Texas Spoofing Test Battery (TEXBAT) dataset. The proposed approach adopts a tracking-level, non-sequential formulation evaluated on static and replay-like spoofing scenarios. Using nine spoof-sensitive features across multiple Pseudo-Random Noise (PRN) codes, the agent classifies clean and spoofed signals into seven classes, achieving 88.1% multi-class accuracy and a macro F1-score of 0.884. The proposed method outperforms SVM, random forest, and autoencoder baselines, demonstrating the effectiveness of reinforcement learning for GNSS spoofing detection. By learning decision policies through interaction and reward feedback, reward shaping improves sensitivity to stealth and replay spoofing without increasing false alarms.