Machine Learning and Statistical Methods for GNSS Spoofing Detection: A Systematic Review and Deep Reinforcement Learning Simulation

Abstract

Global Navigation Systems (GNSS) is essential for global navigation, positioning, and timing across sectors such as transportation, telecommunications, and critical infrastructure. However, the weak, unencrypted nature of GNSS signals makes them highly vulnerable to spoofing attacks, posing serious threats to system reliability and safety. This thesis begins with an introduction to reinforcement learning (RL), providing foundational context on Markov decision processes, model-free and model-based RL, and deep Q-network (DQN) algorithms. Building upon this, it presents a detailed technical review of Global Navigation Satellite System (GNSS) signal architecture, positioning mechanisms, and spoofing vulnerabilities. The widespread reliance on unencrypted, low-power satellite signals renders GNSS systems highly susceptible to spoofing and jamming attacks.

To understand the current state of research, a systematic literature review (SLR) was conducted on AI-based GNSS spoofing and jamming detection strategies. The review identified that while machine learning and deep learning techniques are extensively used, reinforcement learning remains underrepresented. Most prior works focused on binary classification or simulation-based datasets, with limited attention to real-world data.

This thesis also develops and evaluates a multiclass spoofing detection framework using a DQN agent trained on the TEXBAT dataset. Nine spoofing-sensitive features were extracted across eight PRNs per snapshot, forming a 9-dimensional input vector. A stateless OpenAI Gym environment was designed to classify each signal instance as either clean or one of six spoofing scenarios. The model was implemented using Stable-Baselines3 with a Multilayer Perceptron Policy architecture and trained over 80,000 timesteps with ε-greedy exploration and experience replay.

On the test set of 28,391 samples, the model achieved an overall accuracy of 87.6%. Classification accuracy was 90% for ds1, 97% for ds2, 95% for ds3, and 98% for ds4, with F1-scores ranging from 0.922 to 0.979. For more sophisticated attacks, ds7 (sparse stealth) was correctly classified in 81% of cases (F1 = 0.725), and ds8 (replay attack) achieved 71% accuracy (F1 = 0.763). The clean signal class was identified with 81% accuracy, 0.920 precision, and 0.810 recall. These results demonstrate the effectiveness of reinforcement learning in detecting diverse GNSS spoofing attacks using real world tracking features.