Return to Session B3 Next Abstract

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

Closing the Loop: AI-Driven GNSS Interference Detection and Mitigation in Real Time
A. Chamorro, J. Riesco, E. Carbonell, M.P. Bejarano, M.A. Ramirez, A. Gonzalez, GMV
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 8:35 a.m.

Global Navigation Satellite Systems (GNSS) and user algorithms are constantly evolving, allowing user performance to enhance accuracy, availability, and trust in the absolute positioning solutions. This is transversally required by all kinds of high accuracy and safety critical use cases such as autonomous road vehicles, industrial automation, unmanned aerial vehicles (UAVs), precision robotics, and high-accuracy applications such as precision agriculture. In these contexts, GNSS is no longer a “best-effort” sensor; it is part of a safety- and mission-critical chain that must sustain integrity and continuity under highly variable conditions. This shift has also increased the value of GNSS as a target to jeopardize some of the main missions in which GNSS plays a relevant role. Intentional radio-frequency interference, particularly jamming and spoofing—has emerged as a practical, fast-evolving threat capable of degrading service or misleading the navigation solution. Jamming typically reduces availability by overwhelming legitimate satellite signals, whereas spoofing threatens integrity by inducing tracking of counterfeit signals that appear plausible to the receiver. The challenge is amplified by the fact that real operational environments already contain effects that can resemble attacks (e.g., multipath, local RF emissions, ionospheric scintillation), complicating the separation between benign disturbances and malicious behavior.
Traditional monitoring strategies often rely on static thresholds applied to a limited set of indicators. While such approaches can be effective in constrained scenarios, they tend to struggle when operating conditions change rapidly, when multiple interference sources coexist, or when the attacker adapts their strategy. In practice, threshold tuning becomes a continuous compromise between sensitivity and false alarm rate. Addressing these challenges is calling for a new generation of intelligent, adaptable countermeasures capable of learning and responding to emerging threats.
In this work we present an AI-assisted framework for GNSS interference monitoring integrated within GMV GSharp® and its Positioning Engine. The objective is to provide real-time detection and classification of GNSS anomalies that can support operational decisions and mitigation actions within the navigation engine itself. A key innovative step in this framework is the implementation of advanced feature engineering designed to capture the temporal evolution of the original signal features. Unlike traditional threshold-driven options, this AI-driven approach employs data-driven learning to recognize complex patterns across multiple signal observables such as carrier-to-noise density ratio (C/N0), pseudorange residuals, Doppler frequency shifts, and signal quality indicators.
To effectively characterize the evolution of these features, the system calculates new derived features across different temporal windows. Therefore, the input data for training and inference consists of both original features at the current time and the statistical and trend-based features computed for the given temporal windows. Combining these parameters, the model captures multidimensional footprints of interference events, enabling it to better distinguish benign environmental disturbances from deliberate attacks with a higher degree of confidence than conventional implementations.
The framework is trained using a comprehensive labeled dataset built from two complementary sources. First, real recordings of interference scenarios are collected during field trials, showcasing diverse operational conditions and environments to expose the model to realistic attacks. Second, controlled laboratory simulations provide controlled instances of known attack patterns. This labelled dataset allows the model to learn distinctive temporal and spectral characteristics of different interference types. The training process includes iterative optimization and validation to manage the trade-off between missed detection or false alarms—an important consideration in safety-relevant detection, where both can carry operational impact.
During operation, the trained model runs continuously alongside the GMV GSharp® Positioning Engine, ingesting incoming observables and producing a classification of the current situation. The design provides the Position Engine the capacity to act upon these detections: when suspicious behavior or a confirmed attack is detected, the engine can initiate mitigation strategies such as excluding or de-weighting compromised measurements, modifying the fusion strategy, and increasing reliance on complementary sensors (e.g., inertial sensors and/or vehicle ego-motion systems) to preserve navigation continuity. This closed-loop design connects detection to response, reducing time-to-mitigation and supporting resilient operation in applications where short reaction delays can be critical.
Beyond improved detection, the proposed data-driven approach provides practical operational advantages. First, by learning from representative benign conditions, it can reduce false positives triggered by non-malicious effects such as multipath propagation or ionospheric disturbances. Second, it supports evolution over time: as new interference patterns appear, the model can be retrained or updated using additional labeled data, maintaining relevance in a changing threat landscape. Third, classification provides richer situational awareness than binary threshold flags, enabling downstream security logic to consider likely interference type, confidence, and persistence, and tailor mitigation policies accordingly.
Finally, the framework is engineered with deployment constraints in mind. Real-time interference monitoring must deliver timely decisions under limited CPU and memory budgets, and must be compatible with operational GNSS processing pipelines. For this reason, model architectures and feature sets are selected with computational efficiency as a requirement, aiming to provide low-latency inference while maintaining robust performance.
In summary, this work contributes to the advancement of GNSS security by introducing a practical, scalable, and intelligent detection mechanism based on machine learning. Through the combination of robust data collection, algorithmic innovation, and system-level integration, GMV’s AI-enhanced approach offers a powerful tool for ensuring integrity, availability, and reliability in satellite-based navigation. As GNSS continues contributing to the digital infrastructure of the modern world, the adoption of AI-powered defense mechanisms such as the one proposed here will be critical to ensure the next generation of autonomous and precision systems against interference and manipulation.

Return to Session B3 Next Abstract