Ashika Ruwanthi 

Development of Machine Learning Based Models 
for Detecting GNSS Signal Jamming in Real-World 

Scenarios Using AGC Data  

 

 

 

 

 

 

  

Vaasa 2025 

School of Technology and Innovations 
Master’s Thesis 

Sustainable and Autonomous Systems  



2 

 

   

 

UNIVERSITY OF VAASA 
School of Technology and Innovations 

Author: Ashika Ruwanthi 
Title of the thesis:  Development of Machine Learning Based Models for Detecting 

GNSS Signal Jamming in Real-World Scenarios Using AGC Data 
Degree: Master of Computer Science 
Discipline: Sustainable and Autonomous Systems 
Supervisor: Heidi Kuusniemi 
Evaluator Petri Välisuo 
Year: 2025 Pages: 71 

ABSTRACT:  
 
Although Global Navigation Satellite Systems (GNSS) provide high-precision positioning under 
optimal situations, they remain vulnerable to intentional Radio Frequency Interference (RFI) due 
to the low signal power received at ground level and the increasing availability of jamming 
devices. Previous studies have demonstrated that traditional jamming detection methods, such 
as fixed-threshold-based detection, are limited in handling dynamic scenarios and therefore 
remain less efficient under real-world conditions. This study addresses these limitations by 
applying machine learning (ML) techniques to detect GNSS jamming using AGC data. 
 
The primary objective of the study is to analyse supervised and unsupervised machine learning 
(ML) models for detecting jamming using multidimensional features derived from AGC signals. 
The theoretical foundation is based on anomaly detection, statistical learning, and time-series 
signal analysis. Two machine learning models were applied: a supervised classifier (XGBoost) 
and an unsupervised anomaly detector (Isolation Forest). Both were trained and validated on 
AGC data measured during a controlled jamming test in Norway. The experiment is conducted 
under various jamming conditions, which enable realistic and reproducible data collection in a 
dynamic vehicular environment. 
 
Feature extraction was performed using the sliding-window approach across various GNSS 
frequency bands, retaining the temporal and spectral dynamics of AGC. The supervised model 
was trained on labelled samples to distinguish normal and jammed windows, while the 
unsupervised model was trained on normal data only to identify anomalies. The test consisted 
of typical classification metrics and interpretability methods, such as feature importance 
analysis, to understand the model's decisions. 
 
Both models demonstrated the ability to detect jamming effectively. The supervised model 
showed strong performance, revealing that AGC features from specific frequency bands, 
particularly the G1 band, were most indicative of jamming. The unsupervised model performed 
reliably on normal data, although it generated some false predictions, which may be attributed 
to sudden variations in positioning data. Visual analysis offers a deeper understanding of the 
relationship between prediction results and the behavior of positioning parameters. 
 
The proposed ML-based approaches provide better adaptability, eliminate the need for manual 
threshold tuning, and scale effectively across various conditions. They are well-suited for real-
time deployment in GNSS receivers. Future work may enhance these models by combining them 
to provide a total localization solution. 
 

KEYWORDS: (GNSS, Jamming, Detection, Isolation Forest, XGBoost, AGC, Machine Learning) 



3 

 

   

 

Contents  

1 Introduction 8 

1.1 Background and Motivation 8 

1.1.1 Jammertest 11 

1.1.2 Motivation 11 

1.2 Research Problems 13 

1.3 Research Objectives 13 

1.4 Scope of Work 14 

1.5 Thesis Structure 14 

2 Literature Review 16 

2.1 Overview of GNSS 16 

2.2 GNSS Errors and Vulnerabilities for Navigation 18 

2.2.1 Jamming in GNSS Navigation 19 

2.3 GNSS Jamming Detection Methods 21 

2.3.1 Traditional Approaches for Jamming Detection 21 

2.3.2 Machine Learning for GNSS Jamming Detection 23 

2.4 Mitigation Techniques for GNSS Jamming 26 

2.4.1 Traditional Error Mitigation Methods 26 

2.4.2 Machine Learning Based Jamming Mitigation Techniques 29 

2.5 Smartphone Sensors and Sensor Fusion in GNSS Navigation 30 

2.6 Chapter Summary and Research Gaps 31 

3 Methodology 33 

3.1 Research Design and Approach 33 

3.2 Data Collection 33 

3.2.1 Jammer Test Setup 34 

3.2.2 Sensor Data Sources 35 

3.2.3 Data Acquisition 37 

3.2.4 Impact of Jamming for Positioning and AGC Data 39 

3.2.5 The Vehicle Track in Geographical and ENU Coordinates Systems 41 

3.3 Preprocessing and Labelling 41 



4 

 

   

 

3.3.1 Data Preprocessing 41 

3.3.2 Data Labelling for Supervised Machine Learning Model 42 

3.4 Feature Selection and Extraction 43 

3.4.1 Rationale for Using AGC Data 43 

3.4.2 Feature Selection 44 

3.4.3 Feature Extraction 44 

3.5 Machine Learning Model Development 45 

3.5.1 Model Selection 46 

3.5.2 Training and Testing 47 

3.5.3 Tools and Libraries 49 

3.6 Performance and Evaluation Metrics 50 

3.6.1 Evaluation Metrics 50 

3.6.2 Validation Approach 51 

3.7 Summary 52 

4 Analysis and Results 53 

4.1 Overview of the Dataset and Experimental Setup 53 

Dataset Composition 53 

4.2 Performance Evaluation 55 

4.2.1 Supervised Model (XGBoost) 55 

4.2.2 Unsupervised Model (Isolation Forest) 56 

Visual Investigation of IF Results 57 

4.3 Feature Importance for XGBoost Model 59 

4.4 Summary 60 

5 Discussion 61 

6 Conclusions and Future Works 64 

References 65 

Appendices 71 

Appendix A. GNSS Systems Overview with Signal Notation and Frequency 71 

 
 



5 

 

   

 

Figures  
 
Figure 1: Jammertest Locations at Andøya; Red: Bleik; Green: Grunnvatn; Blue: Stave 

(Test Catalogue, 2024) 12 

Figure 2: Trilateration of GNSS Positioning 17 

Figure 3: Illustration of Jamming for GNSS Receivers 19 

Figure 4: The Overall Workflow of the Study 34 

Figure 5: (a) The Test Environment (b) The Geospatial Layout of the Jamming Locations

 35 

Figure 6: GNSS Signal Behaviour for Each Band 36 

Figure 7: The Position and AGC Data over Time for Test Case 2.2.3. 38 

Figure 8: The Jamming Impact on Observed Data 40 

Figure 9: The vehicle track in geographical and ENU coordinates systems 41 

Figure 10: Effect of EMA Smoothing on AGC Data for Each Frequency 42 

Figure 11: AGC Data with Highlighted Normal Period 43 

Figure 12: Visualization of IF Model Predictions and GNSS Positioning Parameters 

During Detected Jamming Events. 58 

Figure 13: Feature importance plot from the trained XGBoost model. 59 

 
 

Tables 
 
Table 1: Summary Table of Traditional GNSS Jamming Detection Methods 22 

Table 2: Comparison of Traditional and Machine Learning-Based Methods for GNSS 

Jamming   Detection 25 

Table 3: An Overview of the Jammer Specifications 36 

Table 4: Dataset Characteristics 54 

Table 5: The Confusion matrix for the XGBoost Model. 55 

Table 6: Classification Metrics for XGBoost Model. 55 

Table 7: The Confusion matrix for the Isolation Forest Model. 57 

Table 8: Classification Metrics for Isolation Forest Model. 57 



6 

 

   

 

Table 9: GNSS Systems Overview with Signal Notation and Frequency (Test Catalogue, 

2024). 71 

 
Abbreviations  
 
AGC – Automatic Gain Control 

AI - Artificial Intelligence 

AV – Autonomous Vehicles 

CNN – Convolutional Neural Network 

CW - Continuous Wave 

DGNSS - Differential GNSS 

DME - Distance Measuring Equipment 

EMA – Exponential Moving Average 

FM - Frequency Modulation  

GAN – Generative Adversarial Network 

GIS - Geospatial Information Systems  

GNSS – Global Navigation Satellite Systems 

GPS Global Positioning System 

IMU – Inertial Measurement Unit 

IF – Isolation Forest 

kNN - K Nearest Neighbours 

KS - Kolmogorov–Smirnov  

LSTM – Long Short-term Memory 

ML – Machine Learning 

PNT - Positioning, Navigation, and Timing  

PPD - Personal Privacy Devices 

PPP - Precise Point Positioning  

RAIM - Receiver Autonomous Integrity Monitoring  

RFI - Radio Frequency Interference 

RNN – Recurrent Neural Network 

RTK – Real-Time Kinematic  



7 

 

   

 

SBAS - Satellite-Based Augmentation System  

SLAM - Simultaneous Localization and Mapping 

SNR - Signal to Noise Ratio 

SVM – Support Vector Machine  

TACAN - Tactical Air Navigation  

UAV – Unmanned Aerial Vehicles 

  
 
  



8 

 

   

 

1 Introduction 

The introduction chapter highlights the growing vulnerabilities of GNSS systems to 

jamming, especially in positioning and navigation. It presents the motivation for using 

real-world data in machine learning based jamming detection using AGC data. The 

chapter also includes research problems, objectives, scope of study, and the structure of 

the thesis. 

 

1.1 Background and Motivation  

Over the last few years, applications of Global Navigation Satellite Systems (GNSS) have 

grown exponentially due to their capability of providing accurate and precise positioning 

and timing information. The global availability of GNSS signals has enabled them to play 

a significant role in numerous industries, including transportation, agriculture, logistics, 

and infrastructure management. As reliance on GNSS continues to increase, particularly 

in high-precision applications such as connected and autonomous vehicles, the 

requirement for real-time and reliable positioning data has become increasingly critical. 

Despite having numerous advantages, GNSS signals are vulnerable to attacks due to the 

low transmission power levels. This inherent vulnerability raises deep concerns about 

the resilience and integrity of GNSS-based services, particularly in safety-critical and real-

time operational environments. 

 
GNSS signals need to propagate nearly 20,000 km to reach the ground as the satellites 

orbit in Medium Earth Orbit (MEO). Normally, the signal strength will decrease over the 

distance. The average signal power decreases to approximately -130 dBm (L1 band) 

when it reaches the GNSS receivers (Spanghero et al., 2025). Therefore, the GNSS signals 

are easily interfered with by radio frequency interference (RFI), which can be either 

intentional or nonintentional. Electronics devices are an example of a non-intentional 

interference source. Even low-power jammers can easily overwhelm the actual satellite 

signals, resulting in poorer positioning performance or complete service denial. 

 



9 

 

   

 

Jamming in GNSS systems refers to the intentional broadcasting of radio frequency 

signals meant to disrupt GNSS receiver operation. The signals are typically broadcast at 

high power and on frequencies near GNSS frequencies, such as L1, L2, and L5, with the 

purpose of malfunctioning the receivers to disrupt their ability to acquire or track 

satellite signals (Radoš et al., 2024). As a result, the receiver may lose satellite 

information or be unable to calculate valid positioning data. Small, inexpensive jamming 

devices can also disrupt GNSS services across extensive ranges, and jamming is a viable 

and significant threat that is easily accessible. Although jamming devices are less 

complex than more advanced threats like spoofing, their extensive use can cause severe 

problems, especially in situations where uninterrupted GNSS access is critical. 

 
GNSS jamming poses a significant threat to a wide range of systems that rely on precise 

positioning, navigation, and timing services. Jamming results in loss of lock, lowered 

accuracy, or GNSS-based failure. GNSS jamming is a considerable topic not only in 

navigation but also in other critical applications such as telecommunications, power grid 

synchronization, aviation, and maritime operations. Cascading failure or operational 

delay can be experienced even for minor interruptions in time-sensitive systems. 

Nowadays, the impact of jamming presents a serious risk to both safety and service 

continuity as the reliance on GNSS applications. 

 
Traditional jamming detection methods in GNSS rely mainly on signal features and 

statistical pattern observation in the receiver. These types of methods are mostly 

categorized into pre-correlation and post-correlation methods. Pre-correlation methods, 

such as monitoring Automatic Gain Control (AGC) levels, statistical signal analysis, and 

transformed domain techniques, are applied at the receiver front-end to detect 

anomalies before signal correlation occurs. Postcorrelation techniques, including the 

analysis of carrier-to-noise density ratio (C/N₀) and correlation function behavior, detect 

interference based on disruptions observed during or after signal acquisition and 

tracking (Reda et al., 2024). Many of these methods depend on predefined thresholds 

or reference models, where deviations in signal power or quality trigger interference 

alerts. While such techniques may work well under controlled or weak interference 



10 

 

   

 

environments, they will fail in strong or complex jamming environments, where signal 

tracking itself may fail (C. Liu et al., 2025). Moreover, prior knowledge of interference 

behavior is required in traditional techniques. Therefore, they are not as appropriate for 

dynamic and uncertain environments. 

 
Traditional detection methods struggle to detect jamming due to the growing complexity 

and sophistication of GNSS jamming techniques. This highlights a need for robust and 

novel detection approaches, such as data-driven machine learning (ML) methods. ML 

algorithms can be trained on vast amounts of historical GNSS data.  This process enables 

the automated extraction of features and the identification of hidden patterns that are 

difficult to detect using traditional threshold-based methods. Compared to static models, 

ML algorithms can learn and adapt to diverse and dynamically changing interference 

scenarios, thereby improving detection accuracy over time. Recent research has 

demonstrated that ML methods are highly effective in detecting and classifying jamming 

signals by leveraging advanced data representations and signal transformations. Their 

flexibility, scalability, and ability to process complex signal characteristics make ML-based 

techniques an essential component of the design of strong and reliable GNSS 

interference detection systems (Reda et al., 2024). 

 
GNSS functionality has been embedded with smartphones, enabling them to capture 

raw measurement data such as pseudorange, Doppler shift, C/N₀, and AGC. Among 

these, AGC is particularly valuable for detecting signal strength variations that may 

indicate the presence of interference. However, despite its relevance, AGC has mostly 

been utilized with C/N₀, rather than as an independent feature in machine learning-

based jamming detection. This study focuses specifically on AGC to explore its potential 

as an independent indicator of GNSS signal integrity. The high availability of smartphones 

offers a cost-effective and scalable platform for interference detection. Thereby, a 

broader geographic coverage and practical deployment in real-world environments can 

be facilitated. By incorporating AGC data into machine learning frameworks, this 

approach aims to develop adaptive and robust jamming detection models that do not 

rely on specialized or high-end GNSS equipment. 



11 

 

   

 

 
Real-world data play a crucial role in developing effective GNSS jamming detection 

methods. Unlike simulated data, real-world data captures the complexity, diversity, and 

randomness of real-world jamming scenarios, which include environmental noise, signal 

variations, and various interference patterns. Training and testing detection models on 

such data ensures better generalization and robustness, enabling the models to perform 

well in real-world settings. This is particularly the case for machine learning approaches, 

which rely on representative data in order to learn meaningful patterns and make sound 

predictions. 

 
1.1.1 Jammertest 

The Jammertest is an event conducted by the Norwegian government that aims to 

promote the robust and intelligent use of Global Navigation Satellite Systems (GNSS). 

The event provides a controlled testbed where experts from industry, academia, and 

public authorities can collaboratively examine GNSS vulnerabilities under realistic 

conditions (Test Catalogue, 2024). It is held annually at Andøya. Unlike traditional 

military jamming exercises, Jammertest is a civilian-accessible environment that focuses 

on testing the effects of jamming, spoofing, and meaconing in a safe and regulated 

setting. 

 
Jammertest features a structured and flexible test catalogue that enables detailed and 

repeatable GNSS interference testing. It supports diverse threat scenarios and promotes 

transparency by publishing machine-readable test data on GitHub, fostering 

collaboration and enhancing global GNSS protection efforts. Further information about 

Jammertest can be found on (Previous Jammertests, n.d.) 

 

1.1.2 Motivation 

The primary motivation for this study is to explore the effectiveness of machine learning 

(ML) models in detecting GNSS jamming using AGC data collected from smartphones. 

Although machine learning-based solutions have been promising in interference 



12 

 

   

 

detection, current research has mostly employed AGC alongside other features such as 

C/N₀ rather than its sole efficacy being assessed. Furthermore, there is a visible lack of 

integration in the comparative study of supervised and unsupervised machine learning 

models in GNSS jamming detection that restricts our knowledge about their relative 

strengths and limitations. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 
This study addresses these gaps by focusing on AGC as an independent input for ML 

models and evaluating both supervised and unsupervised learning techniques. 

Additionally, the research is grounded in real-world data, ensuring practical relevance 

and applicability. This approach supports the development of cost-effective and 

adaptable solutions which are suitable for deployment in dynamic environments.  As the 

reliance on GNSS across numerous sectors, including transportation, emergency services, 

and autonomous technology, continues to grow, it is both opportune and essential to 

develop the resilience of positioning systems to deliberate interference. The objective of 

Figure 1: Jammertest Locations at Andøya; Red: Bleik; Green: 
Grunnvatn; Blue: Stave (Test Catalogue, 2024) 



13 

 

   

 

this study is to further the creation of sophisticated GNSS protection methods through 

the application of a definite and comparative machine learning procedure. 

 

1.2 Research Problems 

While there is an increasing dependence on GNSS for timing and navigation, the systems 

remain highly vulnerable to intentional interference in the form of jamming, especially 

due to the low power of satellite signals at the ground level. Traditional jamming 

detection methods often depend on static thresholds or fixed models, which limit their 

applicability under dynamic, real-world scenarios.  

 
This research addresses the following core problems 

• To what extent can AGC data, collected from consumer-grade GNSS-enabled 

devices, be used as a standalone feature for detecting jamming signals? 

• How can supervised machine learning models be trained on AGC data to reliably 

detect various types of GNSS jamming? 

• What is the potential of unsupervised machine learning techniques to identify 

jamming patterns in AGC data without labelled ground truth? 

• How do supervised and unsupervised models compare in terms of detection 

accuracy, generalizability, and applicability in real-world GNSS jamming 

scenarios? 

 

1.3 Research Objectives 

This research aims to develop an effective and scalable approach for detecting GNSS 

jamming using AGC data alone. The specific objectives are as follows: 

• To examine the viability of AGC as a standalone feature for detecting 

GNSS jamming, using data collected from consumer-grade GNSS-enabled 

devices. 

• To design and implement supervised machine learning models trained 

exclusively on AGC data for the reliable identification of jamming events. 

• To develop unsupervised machine learning models capable of detecting 



14 

 

   

 

jamming anomalies in AGC data without the need for labelled training 

data. 

• To perform a comparative analysis of supervised and unsupervised 

models in terms of detection accuracy, robustness, and adaptability to 

real-world GNSS interference scenarios. 

 

1.4 Scope of Work 

This research is limited to the development and evaluation of machine learning-based 

methods for detecting GNSS jamming using AGC data obtained from smartphones. The 

study focuses exclusively on detecting stationary jamming scenarios. However, it does 

not address spoofing, mobile jamming, or signal mitigation techniques. The 

methodology will involve AGC data preprocessing, the development of both supervised 

and unsupervised machine learning models, and a systematic comparison of their 

performance in terms of accuracy and adaptability. Real-world datasets, specifically from 

the Jammertest event, will be used to ensure the practical relevance of the models under 

realistic signal conditions. The overall goal is to assess the feasibility and effectiveness of 

using AGC data alone for reliable jamming detection, providing a lightweight and scalable 

approach that does not rely on specialized GNSS hardware. 

 

1.5 Thesis Structure 

• Chapter 1 – Introduction 

Introduces the research by outlining the background and motivation. Further, it 

explains the problem statement, research questions, objectives, and scope. 

Finally, the overall structure of the thesis is given. 

• Chapter 2 – Literature Review 

Examines existing studies on GNSS jamming detection, with a focus on traditional 

techniques and the integration of machine learning approaches. It also highlights 

the limited use of AGC data and identifies gaps that this study aims to address. 

 



15 

 

   

 

• Chapter 3 – Methodology 

Describes the research design, including the use of AGC data, data collection 

from real-world sources, preprocessing methods, and the development of 

supervised and unsupervised machine learning models. The chapter also outlines 

the criteria for evaluating model performance. 

• Chapter 4 – Experimental Design and Analysis 

Details the experimental setup, including the definition of stationary jamming 

scenarios, model training and validation processes, and the strategy for 

comparing the performance of different machine learning models. 

• Chapter 5 – Results and Discussion  

Presents the experimental results, evaluates the effectiveness of the supervised 

and unsupervised models, and discusses their practical implications and 

limitations based on real-world jamming conditions. 

• Chapter 6 – Conclusion and Future Work 

Summarizes the key findings, reflects on the research contributions and 

limitations, and proposes directions for future studies, including potential 

expansion to mobile jamming and multi-feature detection frameworks. 

 

 

Consent and Details Regarding AI-Generated Content 

 

I confirm that parts of this master thesis were prepared with the assistance of a Large 

Language Model (LLM) developed by OpenAI, specifically GPT-4o-mini, a variant of the 

GPT-4 architecture. The AI was used as a supportive tool for drafting, phrasing, and 

generating ideas. All final content was critically reviewed and edited by me to ensure 

academic integrity and originality. I acknowledge responsibility for the entire thesis 

content and certify that the use of the AI tool complies with the ethical guidelines and 

policies of my institution. 



16 

 

   

 

2 Literature Review 

This chapter provides an overview of the state-of-the-art research on GNSS signal 

degradation due to jamming, focusing on detection techniques. It briefly discusses 

common sources of error in GNSS, traditional jamming detection techniques, and new 

machine learning techniques. The smartphone-based GNSS data usage for jamming 

detection is also discussed. The chapter ends by highlighting significant areas of research 

deficiency, namely the lack of use of AGC data alone and the lack of comparative analysis 

of supervised and unsupervised machine learning algorithms in real-world scenarios. 

 

2.1 Overview of GNSS 

GNSS is modern satellite navigation systems that provide Positioning, Navigation, and 

Timing (PNT) services worldwide. These include leading global systems like the United 

States' Global Positioning System (GPS), Russia's GLONASS, Europe's Galileo, and China's 

BeiDou. In addition to these, there are regional systems such as Japan's QZSS and India's 

NavIC. These systems are a critical component of a wide range of applications, including 

automotive navigation, aerospace, maritime transport, agriculture, geospatial 

information systems (GIS), emergency response, and others, so in autonomous vehicles 

(G. Li & Geng, 2019). 

 
GNSS positioning relies on the calculation of the time it takes for signals from multiple 

satellites to reach a receiver using distance measurements from at least four satellites. 

As shown in Figure 2, the receiver is able to compute its three-dimensional position and 

corrected time using the trilateration principle (Samalla & Naveen Kumar, 2024).  

 
Over the last few years, GNSS technology has developed significantly in several 

directions. Specht et al., (2020) emphasized that the utilization of multi-constellation 

and multi-frequency function in GNSS navigation. Meanwhile, the GNSS receivers have 

revolutionized in technology that enables miniaturization in size and reducing cost. 

These advancements have helped them to easily integrate with everyday devices such 

as smartphones, vehicles, and IoT systems. The availability of receivers has increases 



17 

 

   

 

with this integration. Additionally, modern positioning techniques like Precise Point 

Positioning (PPP) and Real-Time Kinematic (RTK) have emerged, offering high-precision 

location estimates. Thereby, they are able to ensure centimeter or decimeter level 

accuracy, even when using affordable or mobile hardware (Odolinski et al., 2020). 

 

 

 

 

 

 

 

 

 

 

 

 

 

Atmospheric delays, satellite geometry, multipath propagation, and signal interference 

are the source of errors that can affect for the GNSS receiver performance. Such effects 

are particularly prominent in urban canyons as well as GNSS-denied environments, 

whose signal accuracy and reliability are greatly undermined. 

 
Modern transportation infrastructure, e.g., autonomous cars (AVs) and smart traffic 

systems (ITS), increasingly use GNSS for positioning, routing, and central control of traffic. 

As argued by Ghanbarzadeh et al., (2025), GNSS plays a significant role in maximizing 

efficiency of operations and security. However, as Siemuri et al., (2022) put strong 

emphasis on, susceptibility to interference, particularly purposeful jamming, poses 

enormous risks to such applications and necessitates strong detection and 

countermeasures. 

  

Figure 2: Trilateration of GNSS Positioning 



18 

 

   

 

2.2 GNSS Errors and Vulnerabilities for Navigation 

In vehicular navigation, GNSS is prone to several errors and interferences that may 

reduce the performance of the positioning information obtained. GNSS has become 

beneficial and essential for most modern transportation modes, but this navigation may 

be influenced by a number of factors that cause the receivers to deviate from the actual 

positions, hence affecting not only navigation but also safety, efficiency, and 

functionality of transport systems (Aggrey et al., 2020). These disruptions may stem from 

sources such as ionospheric and tropospheric delays, satellite clock errors, multipath 

propagation, and hardware-related biases, each contributing to varying levels of 

positional inaccuracy depending on environmental and operational conditions (Siemuri 

et al., 2022). 

 
Radio Frequency Interference (RFI) poses a significant risk to the reliability of GNSS 

systems since GNSS signals arriving at the Earth's surface are naturally weak in power 

levels (Mehr & Dovis, 2025). Non-intentional RFI occurs due to emissions from other 

communication systems operating within the vicinity or within the frequency bands of 

GNSS. Dovis, (2015) highlights that typical sources are distance measuring equipment 

(DME), tactical air navigation (TACAN) equipment, or emissions from digital TV 

transmitters and other electronic devices. Such RFIs are typically an artifact of harmonic 

distortion or intermodulation products and are not intended to impact GNSS signal 

performance or likely cause any signal quality degradation or negative impact on 

navigation performance. 

 
Intentional RFI, however, is the purposeful transmission of radio signals for the intent of 

interfering with GNSS operations. This category includes jamming, spoofing, and 

meaconing, which represent levels of increasing sophistication and threat, and jamming 

attempts to overwhelm GNSS receivers with high-power noise signals that make it 

impossible for them to be able to track actual satellite transmissions. Spoofing, on the 

other hand, transmits counterfeit GNSS-like signals in an effort to trick a receiver into 

calculating an incorrect position or time. Mehr & Dovis, (2025) further emphasized that 



19 

 

   

 

meaconing differs somewhat in that it captures and retransmits genuine GNSS signals, 

most often with a delay in time, which provides erroneous positioning information 

without introducing new signals. Such rogue threats are particularly undesirable in 

safety-critical applications and point to the need for successful interference detection 

mechanisms. 

 
2.2.1 Jamming in GNSS Navigation  

Jamming is one of the most critical threats to the integrity and reliability of GNSS signals, 

particularly in safety-critical applications such as vehicular navigation, aviation, and 

maritime systems. It refers to the intentional transmission of high-power radio frequency 

signals intending to disrupt or deny the reception of genuine GNSS signals by 

overwhelming the receiver's front end (Kaplan & Hegarty, 2017). Given that GNSS signals 

are extremely weak when they reach the Earth's surface, typically below the thermal 

noise floor, jammers do not require significant transmission power to be effective over 

short to medium distances. As such, even small, low-cost jamming devices can 

significantly degrade or completely block positioning services. 

The jamming for GNSS receivers is illustrated in Figure 3. 

 

 

 

 

 

 

 

 

 

 

Ferre et al. (2019) provided a comprehensive classification of GNSS jamming signals 

based on their modulation characteristics and complexity. They have described AM 

jammers as simple continuous wave signals, either single- or multi-tone. Chirp jammers 

Figure 3: Illustration of Jamming for GNSS Receivers 



20 

 

   

 

sweep their frequency over time, while FM jammers apply sinusoidal frequency 

modulation. Pulse jammers transmit short, high-power bursts periodically, and 

narrowband noise jammers focus interference within specific spectral bands. Each type 

poses different challenges for GNSS signal detection and reliability. 

 
Jamming signals also vary in complexity, ranging from continuous wave (CW) 

interference and frequency-modulated (FM) chirp signals to advanced, adaptive 

interference. Jamming signals may either be structured as narrowband, focused on 

specific frequency components of GNSS signals, or wideband, which cover entire satellite 

navigation bands  (Elghamrawy et al., 2020). In vehicular contexts, mobile jammers—

often carried by vehicles—can lead to intermittent and geographically distributed signal 

outages, making detection and mitigation particularly challenging (Kreuzer & Munz, 

2021). Furthermore, increased availability of online marketplace jamming devices has 

helped promote the threat. The personal privacy devices (PPDs) are commonly marketed 

as vehicle tracking blockers, even though they are used illegally in most authorities 

(Elghamrawy et al., 2022). 

 
The impact of jamming on GNSS receivers is significant in that it can shut down both the 

acquisition and tracking. In the acquisition process, strong interference in the form of 

jamming can prevent satellite signals from being received by the receiver, whereas in the 

tracking process, jamming can result in loss of lock on the signal. This is then followed 

by erroneous estimates of position, velocity, and time (PVT), or total loss of navigation 

function (Humphreys et al., 2012). This has direct implications in automotive 

applications, where GNSS is usually the primary or alternate source of navigation, 

especially in urban environments where other sensors are already being weakened by 

multipath or signal blockage. In transport networks, such interference may affect traffic 

control, fleet management, and more importantly, autonomous driving systems. 

 



21 

 

   

 

2.3 GNSS Jamming Detection Methods  

GNSS jamming detection is a necessary component in the context of the reliability and 

safety of satellite navigation systems.  These detections are required in situations where 

signal integrity is critical, such as transportation, aviation, and emergency services. 

Unintentional or intentional jamming harms the operation of GNSS receivers by 

introducing high-power signals that interfere with actual satellite signals. These attacks 

can potentially reduce positioning accuracy or cause a general loss of signal. Powerful 

detection mechanisms are essential to enable the implementation of mitigation 

techniques. Various jamming detection methods, ranging from conventional signal 

monitoring techniques to sophisticated, data-driven machine learning approaches, will 

be described under this subsection.  

 

2.3.1 Traditional Approaches for Jamming Detection 

Traditional GNSS jamming detection methods primarily rely on monitoring signal power 

levels, signal-to-noise ratio (SNR) metrics, or other statistical parameters. These 

approaches detect jamming by identifying abnormal behaviors of these parameters, 

which may indicate the presence of interference. However, these methods often depend 

on predefined thresholds. However, these threshold-based methods lead to false alarms 

or missed detections, especially in dynamic or complex environments (Ferre et al., 2019). 

 
Another conventional technique has been mentioned by Radoš et al., (2024) that 

involves analyzing the correlator output of GNSS receivers to detect distortions indicative 

of jamming. Statistical tests, such as the Kolmogorov–Smirnov (KS) test, have been 

applied to monitor these outputs for anomalies. While this method enhances detection 

capabilities, it still faces challenges in accurately identifying jamming signals under 

varying conditions. 

 
Time-frequency analysis methods, including the Wigner–Ville distribution and its 

variants, have been employed to detect GNSS interference (Sun et al., 2016). These 

techniques analyze the energy distribution of signals over time and frequency to identify 



22 

 

   

 

patterns associated with jamming. They are typically afflicted with cross-term 

interference and are computationally intensive, which means that they are not as well-

suited to real-time applications (Sun et al., 2021). 

 
Furthermore, traditional detection methods may not effectively identify sophisticated 

jamming techniques, such as frequency-swept or chirp jammers, which can evade simple 

power or SNR-based detection (Sakorn & Supnithi, 2021a). This limitation underscores 

the need for more advanced detection mechanisms capable of adapting to evolving 

jamming strategies. 

 
The summary of traditional GNSS jamming detection methods is given in Table 1. 

Table 1: Summary Table of Traditional GNSS Jamming Detection Methods 

Detection Method Description Limitation 

Signal power, AGC, and 

SNR monitoring  

(Elghamrawy & 

Noureldin, 2023) 

Detects anomalies in signal 

strength and SNR values 

Prone to false alarms, rely on 

fixed thresholds 

Correlator Output 

Analysis (e.g., KS Test) 

(Zhou et al., 2024) 

Monitors receiver 

correlator outputs for 

distortions 

Sensitivity to environmental 

variations; potential for 

missed detections 

Time-Frequency 

Analysis (e.g., Wigner–

Ville)  

(Sun et al., 2016) 

Analyzes signal energy 

distribution over time and 

frequency 

Computationally intensive; 

affected by cross-term 

interference 

Basic Threshold-Based 

Detection  

(Sakorn & Supnithi, 

2021b) 

Uses predefined thresholds 

to identify interference 

Ineffective against 

sophisticated jamming 

techniques like chirp jammers 



23 

 

   

 

2.3.2 Machine Learning for GNSS Jamming Detection 

Machine learning (ML) techniques have emerged as highly effective tools for GNSS 

jamming detection, offering greater adaptability and accuracy compared to traditional 

signal processing methods. Among supervised learning approaches, algorithms such as 

Support Vector Machines (SVM), k-Nearest Neighbors (kNN), and Decision Trees have 

been widely applied to classify jamming signals. They analyse features like signal-to-

noise ratios, Doppler shifts, and correlation distortions. For instance, Qin & Dovis, (2022) 

utilized the kNN method to effectively detect and classify chirp jamming signals, 

demonstrating the technique’s potential in practical GNSS interference scenarios. SVMs, 

in particular, have demonstrated robust performance in distinguishing between different 

types of jammers, such as continuous-wave and frequency-sweeping signals, by 

leveraging well-structured, labeled datasets. 

 
Recent advancements in deep learning have significantly improved GNSS jamming 

detection, especially by Convolutional Neural Networks (CNNs) and Long Short-Term 

Memory (LSTM) networks. CNNs are effective in automatically learning hierarchical 

spatial features from spectrograms or time-frequency representations of GNSS signals, 

enabling them to identify complex interference patterns that conventional methods 

might miss (C. Liu et al., 2025). Meanwhile, LSTM networks are well-suited for capturing 

temporal dependencies in signal sequences, making them ideal for real-time detection 

of jamming events. In a notable study, Viana et al., (2022) combined CNN and LSTM 

architectures to detect jamming in UAV applications, achieving an accuracy rate of 99%, 

underscoring the high potential of hybrid deep learning models in dynamic 

environments. 

 
Despite their promising performance, DL-based detection methods face several 

operational challenges. A key limitation is the reliance on large volumes of labelled 

training data, which are often scarce in the GNSS domain, especially datasets covering 

varied jamming types, power levels, and environmental conditions. Moreover, deep 

learning models are computationally demanding, which can hinder their deployment on 



24 

 

   

 

devices with limited processing capabilities, such as smartphones or embedded GNSS 

receivers. Another critical issue is the lack of model interpretability; complex neural 

networks often function as “black boxes,” making it difficult to understand the reasoning 

behind their decisions, a drawback that may restrict their use in safety-critical systems 

where transparency is essential (OpenAI, 2025). 

 
Unsupervised learning methods have been explored to address the scarcity of labelled 

data in GNSS jamming detection. Techniques such as clustering and anomaly detection 

can identify unusual patterns in GNSS signals without the need for labelled datasets. For 

example, clustering algorithms have been used to group similar jamming signals, 

facilitating the identification of new or evolving jamming techniques. Anomaly detection 

methods can flag deviations from normal signal behavior, potentially indicating the 

presence of jamming. However, these unsupervised approaches may suffer from higher 

false positive rates and may not always provide clear distinctions between different 

types of jamming. 

 
Hybrid approaches combining supervised and unsupervised learning have also been 

investigated. For instance, semi-supervised learning techniques can leverage a small 

amount of labelled data along with a larger pool of unlabelled data to improve model 

performance. Additionally, self-supervised learning frameworks have been proposed, 

where models are trained to predict parts of the data from other parts, effectively 

learning useful representations without explicit labels. These methods hold promise in 

reducing the dependency on labelled datasets while maintaining high detection accuracy. 

Furthermore, the application of transfer learning has shown promise in GNSS jamming 

detection. By leveraging pre-trained models on large datasets, transfer learning can 

mitigate the challenge of limited labelled data in the GNSS domain. For instance, 

employing a ResNet18 model with transfer learning has achieved high accuracy in 

classifying jamming signals, outperforming traditional ML models. This approach enables 

the adaptation of existing models to new jamming scenarios with reduced training 

requirements, facilitating more efficient deployment in diverse environments (OpenAI, 

2025). 



25 

 

   

 

The Table 2 gives the comparison between traditional detection methods and Machine 

learning based detection of GNSS jamming.  

 

Table 2: Comparison of Traditional and Machine Learning-Based Methods for GNSS Jamming   
Detection 

Feature Traditional methods Machine Learning Techniques 

Detection Approach 

(Radoš et al., 2024) 

Threshold-based 

monitoring of signal 

parameters 

Pattern recognition through 

learned models 

Adaptability 

(Radoš et al., 2024) 

Limited to predefined 

scenarios 

High adaptability to diverse and 

evolving jamming techniques 

Data Requirement 

(Caputo, n.d.) 

Minimal; relies on expert-

defined thresholds 

Requires large, labelled datasets 

for training (supervised); less for 

unsupervised 

Computational 

Demand 

(Ghanbarzadeh et 

al., 2025) 

Low; suitable for real-

time processing on 

limited hardware 

High, may require powerful 

processors or GPUs 

Deployment 

Complexity 

(Ghanbarzadeh et 

al., 2025) 

Simple; easy to 

implement and maintain 
 

Complex; requires expertise in 

ML and continuous model 

updates 

Robustness to 

Attacks 

(Caputo, n.d.) 

Vulnerable to 

sophisticated jamming 

techniques 

Can be robust, but susceptible to 

adversarial attacks 

 

Machine learning–based GNSS jamming detection methods offer a significant 

advancement over traditional approaches by enabling the identification of complex and 

evolving interference patterns through data-driven models. Supervised techniques like 

SVMs and deep learning architectures such as CNNs and LSTMs have shown high 



26 

 

   

 

accuracy in classifying various jamming types. Meanwhile, unsupervised methods, 

including clustering and anomaly detection, help in scenarios with limited labelled data. 

Despite their promise, ML models face challenges related to data availability, 

computational demands, and interpretability. Hybrid and transfer learning approaches 

are emerging to mitigate these issues, making ML a powerful but still evolving solution 

in GNSS jamming detection. 

 

2.4 Mitigation Techniques for GNSS Jamming 

GNSS jamming poses a significant threat to the reliability and integrity of satellite-based 

positioning, particularly in safety-critical applications such as aviation, autonomous 

driving, and military operations. As the reliance on GNSS continues to grow, so does the 

need for effective countermeasures against intentional or unintentional signal 

interference. Mitigation techniques aim to detect, reduce, or eliminate the impact of 

jamming on GNSS receivers, ensuring continuity and accuracy of navigation services. 

These techniques range from traditional signal processing methods, such as antenna 

design and filtering, to advanced data-driven approaches like machine learning, which 

can identify and adapt to interference patterns in real time. This section explores various 

GNSS jamming mitigation strategies, focusing on both conventional and emerging 

solutions to enhance system resilience. 

 
2.4.1 Traditional Error Mitigation Methods 

The accuracy and reliability of GNSS are frequently compromised by various errors, 

including ionospheric delays, satellite clock discrepancies, orbital inaccuracies, and, most 

critically in hostile or urban environments, radio frequency interference such as jamming. 

In response to these challenges, several traditional error mitigation and signal integrity 

techniques have been developed, particularly for use in high-stakes sectors like aviation, 

autonomous vehicle navigation, and military operations (Paziewski et al., 2019). 

 
Among the most widely used techniques is Differential GNSS (DGNSS), which 

significantly improves positioning precision by employing ground-based reference 



27 

 

   

 

stations that track satellite signals and compute real-time correction data. These 

corrections are transmitted to GNSS receivers to filter out common errors such as 

satellite clock drift, orbital errors, and atmospheric delays. DGNSS has been effectively 

applied in marine navigation, geodetic surveying, and aviation (W. Liu et al., 2019). 

However, its effectiveness is inherently limited by the density and distribution of 

reference stations, restricting its utility in sparsely covered regions. 

 
Another key integrity mechanism is Receiver Autonomous Integrity Monitoring (RAIM), 

which operates independently of ground infrastructure. RAIM uses redundant satellite 

measurements and statistical analyses to detect inconsistencies or anomalies in signal 

data (Borhani-Darian et al., 2024). This method is particularly valuable in aviation, where 

navigation errors pose serious safety risks. Nevertheless, RAIM’s performance depends 

on the geometric availability of at least five visible satellites, which can be a significant 

limitation in obstructed environments (Zangenehnejad & Gao, 2021). 

 
Satellite-Based Augmentation Systems (SBAS) offer another layer of correction, 

broadcasting error mitigation data from geostationary satellites. Systems like WAAS in 

the U.S., EGNOS in Europe, and MSAS in Japan provide continent-wide correction 

services, enhancing GNSS accuracy and integrity for applications ranging from civil 

aviation to land transportation (Realini et al., 2017). These systems, however, often 

suffer from high latency and a dependency on terrestrial infrastructure, reducing their 

effectiveness in military or densely built urban areas. 

 
One of the most pressing GNSS threats is intentional jamming, which involves the 

emission of high-powered radio signals that overwhelm the GNSS signal at the receiver. 

As GNSS signals are inherently weak upon reaching the Earth's surface, jammers require 

minimal power to be effective over short ranges (Broumandan & Lachapelle, 2018). In 

response, anti-jamming technologies have been developed, including adaptive antenna 

arrays capable of spatial filtering, notch filters to suppress specific interference bands, 

and dynamic power control to optimize receiver sensitivity (Lachapelle et al., 2018). 

While effective against low-power and narrowband interference, these techniques often 



28 

 

   

 

struggle against high-power, broadband, or adaptive jamming sources, which are 

becoming increasingly prevalent. 

 
Spoofing is a threat that generates fake GNSS signals that can mislead the receiver into 

calculating incorrect positions. It is a distinct and sophisticated threat. Whereas jamming 

directly disrupts signal availability, and it is a quick and brute-force form of attack. 

Despite this distinction, both types of interference highlight the need for more resilient 

GNSS security measures. 

 
To complement these conventional methods, the use of multi-frequency and multi-

constellation GNSS receivers has grown. These receivers can mitigate ionospheric errors 

by comparing signals across different frequencies and reduce dependency on a single 

GNSS system (e.g., GPS) by leveraging signals from Galileo, GLONASS, and BeiDou (Wu 

et al., 2019). This diversification improves robustness, especially in environments where 

line-of-sight is frequently obstructed. 

 
However, the increasing complexity and sophistication of GNSS threats—particularly 

those posed by modern jamming techniques—have exposed the limitations of 

traditional mitigation strategies. Many still rely heavily on fixed infrastructure or static 

algorithms and are not adaptive to evolving attack methods or dynamic environments 

(Chen et al., 2019). Consequently, recent research has begun exploring artificial 

intelligence (AI) and ML approaches to improve GNSS resilience (Dabove & Di Pietra, 

2019). These methods enable systems to learn patterns of interference, predict potential 

errors, and adapt positioning computations in real time (Hu et al., 2023). For example, 

ML algorithms can analyze GNSS signal data to detect jamming attempts early or 

compensate for multipath distortions more effectively than classical filtering techniques. 

 
Ultimately, while traditional techniques such as DGNSS, RAIM, SBAS, and anti-jamming 

filters remain foundational, they are increasingly supplemented by modern approaches 

like AI-based detection, sensor fusion, and high-level cryptography. The integration of 

these tools is essential to secure GNSS navigation, particularly in mission-critical and 

highly dynamic environments. Continued innovation in this field is necessary to ensure 



29 

 

   

 

GNSS can remain a reliable component in future transportation, defense, and 

autonomous systems (Wu et al., 2019). 

 

2.4.2 Machine Learning Based Jamming Mitigation Techniques 

ML has emerged as a powerful tool not only for GNSS error detection but also for 

mitigating signal degradation due to jamming, multipath, and environmental 

interference. Deep learning models such as feedforward neural networks and long short-

term memory (LSTM) networks have shown high potential in learning patterns between 

GNSS signal parameters and their respective error sources (G. Li & Geng, 2019). These 

models are trained using large datasets to distinguish between genuine and altered 

signals, improving location accuracy—especially in complex urban environments where 

multipath effects are prevalent (Borhani-Darian et al., 2024). ML-based suppression of 

delayed signals leads to cleaner, more precise positioning critical for vehicle tracking and 

navigation. 

 
Another significant area of ML application is in sensor fusion, where GNSS data is 

combined with inputs from inertial measurement units (IMUs), LiDAR, and cameras to 

improve accuracy and reliability. Traditional techniques such as Kalman filters have been 

extended by ML models that dynamically adjust sensor weights based on changing 

environmental conditions (Wu et al., 2019). Reinforcement learning and Bayesian 

inference are also being explored to create adaptive systems capable of maintaining 

positional accuracy in GNSS-denied environments like tunnels or dense cities. This is 

particularly relevant for autonomous vehicles and UAVs, where real-time, resilient 

navigation is essential. 

 
ML models are also utilized for predictive error modeling, employing architectures like 

recurrent neural networks (RNNs) and transformer models to capture temporal 

dependencies in GNSS data (Ghanbarzadeh et al., 2025). These systems proactively 

predict and correct GNSS signal anomalies before they affect navigation accuracy. 

Furthermore, anti-jamming and anti-spoofing strategies have been enhanced using 



30 

 

   

 

adversarial learning techniques and generative adversarial networks (GANs), which 

simulate attack conditions and train robust detection models (Wen et al., 2020). As GNSS 

vulnerabilities continue to grow, such countermeasures are essential in securing critical 

navigation systems. 

 
Looking forward, innovations such as quantum machine learning and federated learning 

are anticipated to revolutionize GNSS error mitigation (Szot et al., 2019). Quantum ML 

promises rapid real-time error corrections with enhanced computational power, while 

federated learning enables decentralized model training that ensures user privacy and 

data security. Combined with pattern recognition and unsupervised anomaly detection, 

these technologies will drive the future of secure and highly accurate GNSS systems, 

enabling their reliability even in dynamic, signal-challenging environments (Aggrey et al., 

2020). 

 

2.5 Smartphone Sensors and Sensor Fusion in GNSS Navigation 

Modern smartphones are equipped with a wide array of sensors that significantly 

enhance GNSS-based navigation, particularly in environments where satellite signals are 

weak or unavailable. These sensors—including accelerometers, gyroscopes, 

magnetometers, barometers, and cameras—serve as complementary technologies to 

overcome GNSS limitations in areas like urban canyons, tunnels, or indoors. For example, 

accelerometers and gyroscopes support dead reckoning by tracking linear and angular 

motion when GNSS signals are blocked, although both are prone to drift errors over time 

and require sensor fusion for improved accuracy (Z. Li et al., 2022). 

 
Magnetometers provide heading information and function as digital compasses, proving 

useful in orientation-dependent applications like pedestrian and automotive navigation. 

However, their performance can be affected by electromagnetic interference, especially 

in urban areas, necessitating calibration and integration with other sensor data 

(Robustelli et al., 2019). Barometers offer valuable altitude information, improving 

vertical accuracy in multi-level buildings and supporting use cases like emergency 

response and indoor navigation (Paziewski et al., 2019). Cameras, used in Visual SLAM, 



31 

 

   

 

have emerged as powerful localization tools, enabling the device to construct maps and 

determine positions simultaneously through computer vision and deep learning 

techniques  (Chen et al., 2019). 

 
In the absence of GNSS signals, especially indoors or underground, smartphones 

leverage alternative technologies such as Wi-Fi fingerprinting and Bluetooth beacons. 

These approaches rely on databases of signal characteristics to approximate user 

locations with relatively high precision (Wen et al., 2020). Additionally, inertial sensors 

combined with environmental maps can estimate movement in confined areas like 

subways or tunnels. This redundancy ensures navigation continuity when satellite-based 

systems are insufficient (Robustelli et al., 2019). 

 
To further enhance navigation accuracy, sensor fusion techniques have become central. 

Kalman Filters are widely used to integrate GNSS with IMU data, allowing reliable 

navigation during temporary GNSS outages, such as in urban environments. Particle 

Filters provide advantages in nonlinear conditions by probabilistically estimating position 

based on diverse sensor inputs, making them suitable for indoor navigation scenarios. 

Recently, deep learning-based fusion techniques have gained traction, offering 

adaptability and robustness by learning complex patterns from multi-sensor inputs, 

supporting advanced applications in robotics and autonomous vehicles (Yi et al., 2022). 

 

2.6  Chapter Summary and Research Gaps 

GNSS jamming presents a serious threat to the integrity of navigation services, 

particularly in critical applications such as transportation, aviation, and emergency 

response. Traditional jamming detection techniques, which rely on signal strength, 

correlator outputs, and time-frequency analysis, are often limited by their reliance on 

static thresholds and their inability to detect sophisticated interference methods like 

chirp jamming. In contrast, ML techniques—especially supervised approaches such as 

SVMs and deep learning models like CNNs and LSTMs—offer enhanced adaptability and 

accuracy by learning complex patterns from signal data. Unsupervised and hybrid ML 

approaches also show promise in overcoming the challenge of limited labelled data. 



32 

 

   

 

However, most existing models rely on a combination of multiple signal features and are 

computationally intensive, posing difficulties for real-time deployment on consumer 

devices. 

 
While prior studies have demonstrated the effectiveness of ML-based GNSS jamming 

detection using rich feature sets, little attention has been given to the potential of AGC 

data as a standalone input. Existing work rarely investigates the exclusive use of AGC—

an inherently available and lightweight signal metric, for both supervised and 

unsupervised ML models. Furthermore, there is a lack of comparative studies evaluating 

the performance of these models in real-world, dynamic jamming scenarios using AGC 

data collected from consumer-grade GNSS devices. This research aims to address this 

gap by developing a scalable framework that uses only AGC data for GNSS jamming 

detection, with a focus on accuracy, robustness, and feasibility for real-time applications 

on resource-limited platforms. 

 



33 

 

   

 

3 Methodology 

This chapter outlines the systematic approach employed in developing machine learning 

frameworks for detecting GNSS signal jamming in real-world scenarios using AGC data. 

The methodology is divided into several phases: research design, data collection, 

preprocessing, feature extraction, model development, and evaluation. 

 

3.1 Research Design and Approach 

The research aims to develop both unsupervised and supervised models capable of 

distinguishing between the normal and jammed GNSS signal conditions using AGC data 

collected from the mobile receiver. In addition to model development, the study 

provides a comprehensive comparison between the two approaches. The research 

follows an experiential design, incorporating real-world AGC data and model evaluation 

to ensure the framework captures realistic jamming scenarios and receiver behavior. 

Although the analysis is quantitative in nature, it contributes to a broader understanding 

of jamming detection methodologies. The overall workflow of the study is illustrated in 

Figure 4. 

 

3.2  Data Collection 

The dataset used in this study was collected during the Jammertest conducted in Norway 

under various controlled jamming, spoofing, and meaconing scenarios. The study 

focuses only on the jamming effect on GNSS signals, and the selected scenario involves 

mobile GNSS reception under multiple stationary jammers, representing a realistic and 

challenging detection environment. The vehicle root is an open sky path that promises 

that the other GNSS errors, such as multipath effects, have no effect on the GNSS signals. 

The Figure 5 (a) shows the environment of the path during the test.  

 

 

 



34 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3.2.1 Jammertest Setup 

In the chosen test case, three jammers were deployed at fixed intervals along a 1-

kilometer section of road between Nordmela and Stave. Each jammer was mounted on 

the roof of a stationary vehicle, with antennas oriented in distinct configurations. 

Specifically, all three jammers were positioned such that their antennas pointed upward 

toward a moving test vehicle equipped with GNSS receivers, or antennas were laid flat 

(i.e., facing downward toward the road surface).  

Figure 4: The Overall Workflow of the Study 



35 

 

   

 

The test vehicle is equipped with standard GNSS receivers and mobile phones to collect 

GNSS live data (e.g. ublox positioning, IMU, odometer, AGC). The geospatial layout of the 

jamming locations is illustrated in Figure 5 (b). 

 

 

 

 

 

 

 

 

 

 

 

 

3.2.2 Sensor Data Sources 

Three handheld multi-band jammers, designated as H6.1, H6.2, and H6.3, were used 

during the test. These devices are battery-powered and user-operable via a simple 

interface, including an on/off button, LED indicators, and DIP switches to toggle between 

frequency bands. 

• H6.1: This jammer transmits over six bands, but only two are relevant to GNSS. 

Both fall within the upper L-band (“L1-only”), thereby primarily affecting GPS 

L1/E1/B1C signals. The relevant antenna for GNSS interference is labelled “6”. To 

limit unintentional disruption of non-GNSS signals, only this antenna was 

activated during the test. 

• H6.2 and H6.3: These are also six-band jammers but have broader GNSS coverage. 

They transmit across the L1, L2, and L5 bands, affecting both upper and lower L-

band frequencies. The active antennas for these devices are labelled “4” (L1), “5” 

(L5), and “6” (L2). 

a b 

Figure 5: (a) The Test Environment (b) The Geospatial Layout of the Jamming Locations  



36 

 

   

 

An overview of the jammer specifications, including output power and targeted GNSS 

bands, is provided in Table 3. 

 

Table 3: An Overview of the Jammer Specifications 

Jammer Min Power 

(W) 

Max Power 

(W) 

Test Bands (As per Annex A) 

H6.1 

(Jammer 3) 

0.631 0.631 G1, L1, E1, B1C 

H6.2 

(Jammer 2) 

0.3981 1.000 
L1, E1, B1C, E6, B3I, G2, L2, E5b, 

B2b, B2I, G3, L5, E5a, B2a 
 

H6.3 

(Jammer 1) 

0.3981 1.000 
L1, E1, B1C, E6, B3I, G2, L2, E5b, 

B2b, B2I, G3, L5, E5a, B2a 
 

 

Jammer 1 and 2 can be categorized as broadband jammers because they affect all the 

bands. However, Jammer 3 can be identified as a narrowband jammer that affected only 

a single frequency band (L1). GNSS signal behaviour for each frequency band is shown 

in Figure 6. 

 

 

 

 

 

 

 

 

 

 

  Figure 6: GNSS Signal Behaviour for Each Band 



37 

 

   

 

Based on the figure above, Jammers 1 and 2 appear to have impacted all observed 

frequency bands. In contrast, Jammer 3 primarily affected the 1575.42 MHz band and 

other frequencies have not responded to the Jammer 3.  

 

3.2.3 Data Acquisition  

The AGC data used in this study was collected using the GNSS Logger application installed 

on a Google Pixel smartphone. This application has captured raw positioning data, IMU 

data, and AGC data during the test period. The test data is stored in an HDF5 file, and 

the details about each test case have been stored in a separate metadata file. For the 

current analysis, the AGC data and positioning data are extracted and processed. 

However, positioning data is used only for the data labelling process.  

 

AGC Data 

AGC mechanism is significant in GNSS signal jamming detection because it dynamically 

adjusts the gain value of the receiver to maintain the same signal strength even with 

fluctuation of input power. As RFI, i.e., intentional jamming, is introduced in the GNSS 

band, thermal noise floor rises, and AGC consequently decreases gain in a bid to prevent 

signal saturation (Spens et al., 2022) The decrease reveals as a measurable reduction of 

AGC values, which itself can be utilized as a consistent metric of abnormal signal 

conditions. Because AGC reacts with an instant response to changes in the input's power 

within a specific range of frequencies, stable changes of its output, particularly steep 

decreases, can signify the presence and relative power of the jamming signals. 

 

Figure 7 presents the AGC test data and positioning information collected during Test 

2.2.3, in which the vehicle was driven toward Nordmela. The first three plots display the 

vehicle's geographical coordinates, north, east, and height. The fourth plot illustrates the 

quality factor (Q), an indicator of the reliability of satellite observations, and the number 

of visible satellites is in the fifth plot. The next plot is the velocity variations over time. 

Notably, data gaps appear in these plots as the vehicle passes near the jamming sources.  

 



38 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 
Figure 7: The Position and AGC Data over Time for Test Case 2.2.3. 

Ja
m

m
er

 1
 

Ja
m

m
er

 2
 

Ja
m

m
er

 3
 



39 

 

   

 

The final graph shows variations in the AGC values over time. A sharp variation can be 

noted during the first two jamming periods, while the third jamming period does not 

show a significant effect on AGC data. Further, the impact on each band is also varied. 

 

It can be observed that Jammer 1 and Jammer 2 significantly disrupted the positioning 

system, leading to a loss or degradation of the computed positional data. In contrast, 

Jammer 3 had no discernible impact on the integrity of the positioning information. 

 

3.2.4 Impact of Jamming for Positioning and AGC Data 

The impact of jamming on positioning accuracy and AGC data during the first jamming 

event is illustrated in Figure 8. A sudden loss of east and north values occurs as the 

vehicle approaches the jammer, and it takes approximately 20 seconds for the receiver 

to reacquire valid positioning data. The position values decrease smoothly during the 

normal (non-jammed) period. However, it is possible to observe a jump in height just 

before data loss, while there are minor variations under normal conditions; The quality 

factor is 5 under normal conditions, representing a single-point positioning solution.  

This parameter also drops as the vehicle enters the jamming zone. 

 
As shown in the fifth plot, the number of visible satellites gradually decreases, and the 

jammer disrupts satellite visibility for more than one minute. A sudden spike in velocity 

is also observed just before data loss. Overall, the positioning data is unavailable for 

approximately 50 seconds during the jamming period, and this duration may vary from 

jammer to jammer depending on their signal strength. 

 
The AGC data shows a gradual slope lasting about 25 seconds before falling sharply as 

the vehicle passes the jammer. The fall continues for approximately 10 seconds, after 

which the recovery in AGC values starts, returning to normal after a short delay. It can 

also be seen that jamming effects are not the same in all GNSS frequency bands. 

 

 



40 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 Figure 8: The Jamming Impact on Observed Data 



41 

 

   

 

3.2.5 The Vehicle Track in Geographical and ENU Coordinates Systems 

The localization of the test vehicle in geographical and ENU coordinates systems are 

shown in Figure 9.  

 

 

 

 

 

 

 

 

 

 

3.3 Preprocessing and Labelling 

3.3.1 Data Preprocessing 

To enhance signal quality and reduce the impact of high-frequency noise, Exponential 

Moving Average (EMA) smoothing was applied to the AGC data before model training. 

EMA is a type of low-pass filter that assigns exponentially decreasing weights to older 

observations. This helps in capturing short-term trends effectively while preserving the 

overall shape of the signal (OpenAI, 2025). 

 
Compared to simple moving averages, EMA responds more quickly to recent changes in 

the data, which is beneficial in detecting rapid changes caused by jamming events. By 

applying EMA smoothing, fluctuations due to random measurement noise are 

attenuated, allowing the machine learning models to focus on more meaningful 

variations in the AGC signal that are indicative of interference or normal conditions. 

 

 

 

Figure 9: The vehicle track in geographical and ENU coordinates systems 



42 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 
Figure 10 demonstrates how EMA smoothing enhances the AGC signal by reducing short-

term fluctuations while preserving overall trends. The figure shows that the smoothed 

signals for each frequency band follow the same pattern as the original signals while 

reducing the unwanted spikes of the signal.  It would be very useful to detect anomalies 

in the AGC data.  

 
3.3.2 Data Labelling for Supervised Machine Learning Model 

This study examines supervised and unsupervised machine learning algorithms for 

detecting jamming in GNSS. One of the biggest challenges is the absence of an explicit 

ground truth label for the data. The challenge has been addressed by using positioning 

data as a reference for labelling. Time slots during which positioning data, including 

latitude, longitude, quality indicator (Q), and number of visible satellites, are absent, are 

considered a potential jamming event. By contrast, data points containing valid 

positioning information are considered normal conditions. The AGC data in jamming 

period is labelled as ‘1’, and normal data are taken as ‘0’. 

 

Figure 10: Effect of EMA Smoothing on AGC Data for Each Frequency 



43 

 

   

 

Figure 11 illustrates how the labelling process captures the jamming and normal period. 

The shaded green areas represent the normal period. The third jamming period is not 

taken as jamming since it does not affect the positioning.  

 

 

 

 

 

 

 

 

 

 

 

 

 

3.4 Feature Selection and Extraction 

Feature engineering is one of the most important steps in developing consistent machine 

learning models for GNSS signal anomaly detection. In this work, I focused on extracting 

temporal and statistical features from AGC data, a signal power metric, to detect the 

variations that are indicative of jamming. The same pipeline was used for supervised and 

unsupervised learning models. 

 

3.4.1 Rationale for Using AGC Data 

AGC values are actively controlled by GNSS receivers to maintain constant signal power. 

In normal operation conditions, AGC values should be within some expected range. 

However, in jamming conditions, artificial signal interference causes abnormal changes 

in signal power and results in prominent AGC adjustments. Therefore, fluctuations in 

AGC values are a robust method to determine probable interference, i.e., jamming. 

Figure 11: AGC Data with Highlighted Normal Period 



44 

 

   

 

3.4.2 Feature Selection 

To effectively identify GNSS signal anomalies such as jamming, an underlying dynamics-

based feature selection of the AGC signal is essential to ensure successful detection. For 

this study, domain expertise and empirical observation of the AGC during clear and 

jamming conditions guided feature selection. Since the values of the AGC respond to 

changes in the received signal power, features representing central tendency, extreme 

measurements, and temporal variability were identified as the most informative. 

 
The selected features include: 

• Mean: to capture the central value of the AGC in a window. 

• Minimum and Maximum: to identify spike values potentially due to sudden 

variation of the signal. 

• Range (max - min): as a measure of short-term signal volatility. 

 
Other aspects, such as slope and standard deviation, were also explored, but were 

excluded from the final model upon initial performance evaluation and to reduce 

dimensionality, especially within the supervised environment. The selection is made for 

balancing computation with detection performance, especially when employed in the 

sliding-window method. 

 
3.4.3 Feature Extraction 

For feature capture extraction that indicates local deviations in the AGC signal typical of 

interference, a sliding window feature extraction was applied. The feature extraction has 

the capability to observe signal features over time locally by dividing the AGC data into 

overlapping windows of the same size (window size) and computing descriptive statistics 

for each window. The temporal resolution and overlap of the output feature set are 

governed by the sliding step (step size). The selection of window size and step size needs 

to be done carefully. The window size should be large enough to capture the meaningful 

features. However, it should be kept as small as possible to ensure that the window does 

not lag behind real data (G. Huang et al., 2022). 



45 

 

   

 

The AGC signal is captured in four unique frequency bands, which are G1, G2, G3, and 

G4 (1575.42 MHz, 1176.45 MHz, 1602.00 MHz, and 1561.096 MHz, respectively). They 

correspond to different GNSS signal carrier frequencies, and the AGC response is varied 

depending on the interference affecting each band. To preserve the separate temporal 

dynamics of each band, features were extracted separately for each AGC signal, creating 

a multidimensional feature space. For each frequency band in a given window, mean, 

maximum and minimum, and range were computed. 

 
This results in a feature vector that includes the above metrics for each of the four bands. 

For example, with three features per band across four bands, each window yields a 12-

dimensional feature vector (e.g., G1_mean, G1_min, G1_max, ..., G4_max). 

 
In supervised learning scenarios, a label is also assigned to each window using the 

majority class (e.g., normal or jamming) present among the samples in that window. This 

allows the features to be used for both classification and unsupervised anomaly 

detection. 

 
The multi-dimensional feature set retains the band-specific characteristics of the AGC 

signal, which is critical for distinguishing between normal variability and targeted 

interference affecting one or more frequency bands. 

 

3.5 Machine Learning Model Development 

This section outlines the machine learning models developed for GNSS jamming 

detection, covering both a supervised learning approach using Random Forest and an 

unsupervised method based on Isolation Forest. Each model was selected based on its 

suitability for the data characteristics and the nature of the detection problem. 

 

 

 



46 

 

   

 

3.5.1 Model Selection 

To address the issue of GNSS signal anomaly detection, unsupervised and supervised 

learning approaches were employed to manage different labelling data situations. Two 

models were selected for their use in high-dimensional time-series features, noise 

resistance, and proven performance in anomaly detection issues: 

 
Isolation Forest (Unsupervised) 

For unsupervised anomaly detection, the Isolation Forest model was selected. The 

model is highly suitable for identifying infrequent and anomalous patterns in high-

dimensional space without requiring ground truth. It operates by constructing random 

decision trees and isolating observations; anomalies tend to be easier to isolate and 

therefore have shorter average path lengths. 

 
Isolation Forest was chosen due to the following advantages: 

• No requirement for labelled data, so usable in early-stage or real-time detection 

where ground truth is absent. 

• Efficient and scalable, applicable to large time-series datasets generated by AGC 

signals. 

• Effective for high-dimensional feature sets, such as those derived from multiple 

frequency bands. 

In this study, the model was trained on features extracted from presumed “normal” 

periods, and anomalies were detected during test windows by comparing their isolation 

depth against the learned distribution. 

 
XGBoost (Supervised) 

For the supervised learning scenario, the Extreme Gradient Boosting (XGBoost) 

algorithm was employed. XGBoost can be defined as a tree-based ensemble algorithm 

that builds additive models in a forward stage-wise manner and optimizes a 

differentiable loss function. Strong prediction performance, regularization, and 

management of class imbalance are the main advantages of the model. 



47 

 

   

 

XGBoost was chosen due to: 

• High classification accuracy, especially with structured tabular features. 

• Integrated missing value and outlier treatment, which is useful when exploring 

GNSS signal environments. 

• Interpretability, via feature importance scores, to explore which AGC bands and 

metrics are most predictive of jamming. 

 
The model was trained using the labelled data with the feature windows labelled as 

normal or jammed, depending on the ground truth derived from controlled jamming 

experiments. The XGBoost classifier gives a probability of jamming, which can be 

thresholded to generate binary predictions. 

 
3.5.2 Training and Testing 

The training and evaluation procedures were tailored for both supervised and 

unsupervised learning models, reflecting their respective assumptions about the 

availability of labelled data. Both models utilized a shared feature extraction process 

from a sliding window algorithm on AGC data. 

 
Supervised Model: XGBoost Classifier 

Training and testing of the XGBoost model were a standard supervised machine learning 

pipeline: 

 
Feature Preparation 

As outlined in Section 3.4, a sliding window technique was used to compute statistical 

features—including mean, minimum, maximum, range, and optionally slope—from each 

AGC frequency band. Each window received a binary label (jamming or normal) based 

on the most occurring ground truth value in the window. 

 
Data Splitting 

The dataset was divided into training and testing subsets using an 80/20 stratified split 

to preserve the class distribution, which is especially important in the case of class 



48 

 

   

 

imbalance.  Stratification served to guarantee that both the normal and jamming cases 

were adequately represented in each subset. 

 
Model Training 

The XGBoost model was trained using binary logistic loss and default hyperparameters. 

This gradient-boosted decision tree algorithm is specifically suited for tabular data with 

a mix of distributions and was selected by virtue of high accuracy and resistance to 

overfitting. 

 
Model Evaluation 

The classifier was evaluated on the test set on the basis of accuracy, precision, recall, and 

F1-score, providing an overall indication of whether or not it can detect jamming 

instances correctly. 

 
Unsupervised Model: Isolation Forest 

The Isolation Forest model was trained on normal (not jamming) data alone and is 

hence particularly well-suited for anomaly detection scenarios where labelled anomaly 

data may not be easily available. 

 
Feature Preparation 

The same sliding window approach and statistical feature extraction were applied, as in 

the supervised case, to ensure consistency in feature representation. 

 
Training and Testing Strategy 

A known jamming interval was predefined based on the experimental setup, and all 

data outside this interval was considered normal. This normal subset was used to train 

the Isolation Forest, which learns the structure of typical AGC behavior by isolating 

observations through random partitioning. 

 
 
 
 
 



49 

 

   

 

Data Augmentation and Scaling 

To enhance the model’s generalization to unseen normal variations, data augmentation 

techniques were applied to the training set: 

• Jittering introduced small random noise. 

• Scaling altered the amplitude of features to simulate dynamic signal 

environments. 

• All features were standardized using StandardScaler prior to model fitting. 

 

Anomaly Detection 

After training, the model was applied to the full dataset. Any feature window with a 

significantly short average path length (i.e., easy to isolate) was flagged as an anomaly. 

These were interpreted as potential jamming events. 

 

3.5.3 Tools and Libraries 

All data analysis, processing, and modeling were conducted using the Python 

programming language, thanks to its extensive data science and machine learning 

ecosystem. Google Colab provided a flexible platform to develop the models. The 

following frameworks and libraries were used throughout the study: 

 

Scikit-learn 

Used for both the development of the Isolation Forest and XGBoost models (via the 

sklearn.ensemble and xgboost interfaces, respectively) as well as feature scaling 

(StandardScaler), splitting data, and metrics for measuring performance. 

 
Pandas and NumPy 

Heavily utilized for data manipulation, pre-processing, and numerical computation. 

Pandas provided simple-to-use functionality for time-indexed AGC data manipulation 

and label sliding window manipulation, and NumPy provided support for high-speed 

numerical computations in feature extraction and augmentation. 

 



50 

 

   

 

Matplotlib and Seaborn 

Employed to visualize AGC signal behavior, model predictions, and feature distributions. 

Further, the raw data visualization was done using these libraries. These libraries 

facilitated the generation of high-quality plots used for both exploratory analysis and 

result presentation. 

 
Custom Python Functions 

Several domain-specific functions were developed to support: 

• Feature extraction from AGC signals using sliding windows. 

• EMA smoothing to denoise raw AGC measurements. 

• Data augmentation techniques (jittering and scaling) to enrich the training data 

for unsupervised learning. 

 
These tools collectively enabled an efficient, reproducible, and scalable workflow for 

GNSS signal anomaly detection. 

 

3.6 Performance and Evaluation Metrics 

Quantitative comparison of the performance of supervised and unsupervised models 

was conducted through standard classification metrics. These metrics provided a 

comprehensive evaluation of how each model effectively detects GNSS signal anomalies 

and, more crucially, jamming events in various signal conditions. 

 
3.6.1 Evaluation Metrics 

The following performance indicators were employed: 

• Accuracy: The proportion of correctly classified instances (both normal and 

jamming) over the total number of predictions. 

• Precision: The proportion of correctly identified jamming instances out of all 

instances predicted as jamming (i.e., the ability to avoid false alarms). 

• Recall: The proportion of actual jamming occurrences correctly identified, which 

means the ability to recognize actual jamming. 



51 

 

   

 

• F1-Score: Harmonic mean between precision and recall, providing a balanced 

measure, particularly useful in class-imbalanced situations. 

• Confusion Matrix: Tabular summary reporting true positives, true negatives, false 

positives, and false negatives. 

 
These values were calculated with the in-built functions of the scikit-learn library. The 

classification report and confusion matrix provided information regarding model 

strengths and failure modes. 

 
3.6.2 Validation Approach 

For the supervised model (XGBoost), performance was evaluated using 3-fold cross-

validation on the training set to ensure robustness and reduce variance in the results. 

Final evaluation was performed on a held-out test set following cross-validation. 

For the unsupervised model (Isolation Forest), normal (non-jammed) data alone was 

trained upon. Testing was performed on the full dataset, including known jamming 

intervals. For aligning predictions to ground truth labels: 

 

• The prediction DataFrame (full_data) and the feature-label DataFrame 

(features_df) were merged based on timestamps to ensure alignment. 

• Only matching entries were retained using an inner join. 

• True labels and predicted labels were extracted and used to compute the 

confusion matrix, classification report, and accuracy. 

 
This evaluation procedure ensured consistency and allowed for direct comparison 

between the predicted anomalies and the actual jamming events annotated in the 

dataset. 

 

 

 



52 

 

   

 

3.7 Summary 

In this chapter, the methodology for the detection of GNSS jamming using supervised 

and unsupervised machine learning was presented. The process began with the 

preprocessing of AGC data, e.g., frequency-specific exponential moving average (EMA) 

smoothing, followed by feature engineering using a sliding window strategy. Key 

temporal and statistical features were extracted independently from four frequency 

bands, resulting in a multidimensional representation of signal behavior suitable for 

anomaly detection. 

 
Two models were used: Isolation Forest in unsupervised and XGBoost for supervised 

classification. Both models underwent tailored training protocols. The Isolation Forest 

was trained on normal signal patterns alone, with jittering and scaling augmentation 

incorporated to improve generalization. The XGBoost model was trained on labelled data 

with a stratified train-test split and validated by 3-fold cross-validation to check the 

robustness. 

 

Standard accuracy, precision, recall, and F1-score performance measures were utilized 

for model evaluation. Results were obtained by cross-referencing prediction outputs 

with ground truth labels using timestamp-based merging to enable the generation of a 

confusion matrix and classification report, and further thorough performance analysis. 

 
Overall, this methodology provides a scalable and effective framework for GNSS signal 

anomaly detection, with potential applicability to real-time and embedded systems. 



53 

 

   

 

4 Analysis and Results 

This chapter provides a detailed analysis of the performance of supervised and 

unsupervised ML models. It presents an overview of the experimental data set and the 

results obtained from each model. The chapter also includes the validation and feature 

importance.  

 

4.1 Overview of the Dataset and Experimental Setup 

This section outlines the dataset used for model training and evaluation. The data 

consists of AGC signal measurements collected from a smartphone GNSS receiver.  The 

data are labelled to identify periods of normal operation and jamming. These labels 

served as the ground truth for the supervised learning model. The dataset was 

preprocessed using EMA smoothing to remove unwanted spikes, and feature extraction 

through sliding windows was employed to prepare the data for both supervised and 

unsupervised modeling. Additionally, data augmentation was used across the board to 

enhance the data. 

 
Dataset Composition 

The dataset used here was constructed from AGC measurements in a real-world 

vehicular GNSS environment. The sliding window approach was used to capture the 

dynamic character of AGC as it changes with varying signal conditions, as follows: 

• Window size: 2 samples  

• Step size: 1 sample (50% overlap) 

 
Each window represents a short-term temporal snapshot of AGC behavior across four 

GNSS frequency bands. The increase in window size may lead to lag the window behind 

the real data, meaning that the jamming prediction occurs after the vehicle reaches its 

actual location. A large window size are not suitable for jamming detection. This process 

generated a total of 2846 labelled samples, used for both supervised and unsupervised 

learning models. 



54 

 

   

 

The dataset was annotated based on known time intervals of jamming, yielding a binary 

classification task: 

• Label 0 – Normal: Signal windows without any known jamming interference. 

• Label 1 – Jammed: Signal windows overlapping with confirmed jamming events. 

 
From each window, a set of statistical features was extracted per frequency band. These 

include the mean, minimum, maximum, and range, which collectively capture both 

signal amplitude and variation trends over time. The resulting feature space is 

multidimensional, reflecting the behavior of the AGC across each of the four frequency 

bands independently. This allows the models to leverage frequency-specific signal 

behavior in order to better detect anomalies. 

 

Table 4: Dataset Characteristics  

Property XGBoost Isolation Forest 

Total number of samples 2846 2846 

Number of Training Samples 80% of total samples Normal data 

Number of testing samples 20% of total samples All data 

Sampling Frequency 8 Hz 8 Hz 

Duration of the dataset 15 min 52 sec 15 min 52 sec 

Window size  2  2  

Step size 1 (50% overlap) 1 (50% overlap) 

Feature extracted per window Mean, Min, Max, Range Mean, Min, Max, Range 

Total features per sample 16 (4 x 4) 16 (4 x 4) 

Label for Training Yes No 

 

 

 

 

  



55 

 

   

 

4.2 Performance Evaluation 

The detailed assessment of both models, which are trained for jamming detection in 

GNSS, is discussed in this section. Performance was tested out using a combination of 

classification evaluation metrics, including accuracy, precision, recall, F1-score, and 

confusion matrices, along with visual checks of predicted vs. actual labels for Isolation 

Forest. Additionally, 3-fold cross-validation was also applied for the supervised model to 

evaluate model robustness. 

 
4.2.1 Supervised Model (XGBoost) 

The XGBoost model was trained using labelled AGC features extracted as detailed in 

Section 3.4. The dataset was split into 80% training and 20% test subsets. In addition, 3-

fold cross-validation was done during training to ensure model stability and reduce the 

risk of overfitting. 

 
The confusion matrix for the XGBoost model is given in the Table 5. 

Table 5: The Confusion matrix for the XGBoost Model. 

 Predicted Normal Predicted Jammed 

Actual Normal 514 1 

Actual Jammed 1 54 

 

The classifier achieved an overall accuracy of 99.65%, correctly identifying most 

instances of both normal and jammed conditions. 

 

Table 6: Classification Metrics for XGBoost Model. 

Class Precision Recall F1-Score 

Normal (0) 1.00 1.00 1.00 

Jammed (1) 0.98 0.98 0.98 

 



56 

 

   

 

Moreover, the mean cross-validation score is 98.21%, which indicates that the model is 

robust for overfitting.  

 
Interpretation 

• High Recall (0.98): The model successfully detects 94% of actual jamming events, 

which is crucial for maintaining GNSS navigation security. 

• High Precision (1.00): Few normal signals are mistakenly classified as jamming, 

indicating a low false positive rate. 

• Strong F1-Score (0.98): Shows a well-balanced trade-off between precision and 

recall for the jammed class. 

• Robust Performance: Despite class imbalance, the model performs well on both 

classes, with particularly strong generalization to the minority (jammed) class. 

 
4.2.2 Unsupervised Model (Isolation Forest) 

For the unsupervised scenario, the Isolation Forest model was trained using only 

normal-labelled samples, in a real-world anomaly detection application where jamming 

patterns are unknown at training time. labelled samples were used for evaluating the 

model, not for training purposes. 

 
Data Handling Strategy: 

• Only normal data samples were used for training, as the model needs to detect 

outliers. 

• Jammed windows were added in the test set to evaluate anomaly detection 

performance. 

• Feature standardization (with StandardScaler) and data augmentation (jittering 

and scaling) were employed to introduce robustness. 

 
The confusion matrix and Classification matrix for the Isolation Forest model are given 

in the Table 7 and Table 8 respectively. 



57 

 

   

 

Table 7: The Confusion matrix for the Isolation Forest Model. 

 Predicted Normal Predicted Jammed 

Actual Normal 2548 24 

Actual Jammed 9 265 

 

Table 8: Classification Metrics for Isolation Forest Model. 

Class Precision Recall F1-Score 

Normal (0) 1.00 0.99 0.99 

Jammed (1) 0.92 0.94 0.94 

 

Although they were trained on unlabelled jamming images, Isolation Forest accurately 

detected unusual patterns of AGC behavior that overlapped with well-known periods of 

jamming. The relatively good recall indicates good sensitivity but at the cost of some 

false positives (true, regular samples incorrectly identified as anomalies)—a trade-off to 

be expected in unsupervised anomaly detection. The contamination value needs to be 

selected carefully to get better performance.  

 
Visual Investigation of IF Results 

The visual representation of the IF model predictions and the positioning data during 

the jamming period are shown in Figure 12. The topmost plot shows the AGC data with 

predicted jamming points (red ‘x’), and the jamming windows based on the model 

predictions are indicated by shaded time intervals. The other plots correspond to 

positioning data, including latitude, longitude, height, quality factor, number of 

satellites, and velocity. The results highlight that the shaded period is bigger than the 

period in which the positioning data is lost. However, the model has captured the 

sudden variations of height, number of satellites, and velocity. 

 

 

 

 



58 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 12: Visualization of IF Model Predictions and GNSS Positioning Parameters During 
Detected Jamming Events. 



59 

 

   

 

4.3 Feature Importance for XGBoost Model 

The supervised XGBoost classifier provides intrinsic feature importance scores, allowing 

us to identify which AGC-based features contribute the most to the classification task. 

The relative importance of each feature from different AGC signal bands (G1–G4) is 

shown in the bar chart of Figure 13. 

 

 

 

 

 

 

 

 

 

 

 

 

 

Key Observations 

• Dominant Features: 

The two most influential features were G1_min and G1_mean, contributing 

approximately 48% and 43%, respectively, to the model’s decision-making 

process. This suggests that minimum and average AGC values from the G1 band 

are highly indicative of jamming events, likely due to G1's sensitivity to power 

disruptions. 

• Moderate Importance: 

G1_max and G3_mean exhibited marginal importance, indicating that other 

bands contribute to the classification, albeit less significantly. 

 

 

Figure 13: Feature importance plot from the trained XGBoost model.  



60 

 

   

 

• Low-Impact Features: 

Most features derived from bands G2, G3, and G4—including range, max, and 

mean—had negligible importance scores. This implies that AGC anomalies 

caused by jamming predominantly manifest in the G1 band, which could relate 

to the receiver’s frequency handling or antenna configuration. 

 
Implications for Model Simplification 

These results suggest that a dimensionality reduction approach may be viable without 

significant performance loss. For instance, feature engineering could prioritize G1 

statistics in future iterations, reducing complexity and improving interpretability. 

 

4.4 Summary 

This chapter provided a comprehensive evaluation of machine learning models for GNSS 

jamming detection based on AGC signal features. The dataset of 2846 time-window 

samples labelled by a sliding window method was utilized as the basis for training and 

testing supervised and unsupervised models. 

 
The XGBoost classifier under supervision worked extremely well with respect to 

prediction accuracy of 99.6%, precision of 0.98, recall of 0.98, and an F1-score of 0.98 

for jamming detection. Analysis of feature importance revealed that G1_min and 

G1_mean statistics of AGC from the G1 frequency band were the most significant in 

performing the classification. 

 
The Isolation Forest model (unsupervised) also distinguished well jamming conditions 

from normal conditions at a rate of accuracy of 98% and balanced the detection of 

anomalies when trained solely on normal data. The model has captured not only the 

position data loss period but also the sudden variation of some positioning data.  

These findings verify the efficacy of using AGC-based features in jamming detection and 

certify both learning paradigms to function effectively in practical signal environments. 

The comparative examination of model performance and deployment implications are 

discussed in the next chapter. 



61 

 

   

 

5 Discussion 

This chapter discusses the results obtained from both supervised and unsupervised 

models and key findings. It also includes the discussion of results over previous works and 

applicability of models.  

 
Supervised and unsupervised machine learning models were developed for GNSS 

jamming detection using AGC data in this study. The supervised XGBoost model has 

achieved a high classification accuracy of 99.65% which indicates that the model has 

correctly classified nearly all samples. The precision and recall scores of 0.98 for the 

jammed class demonstrate the model’s strong ability to accurately detect jamming 

conditions. The confusion matrix revealed only two misclassifications out of over 500 

samples, showing strong reliability in differentiating between normal and jammed 

conditions. The model also demonstrated good generalization even with minority 

jamming samples, achieving a cross-validation mean accuracy of 98.21%, indicating 

robustness and minimal overfitting. 

 
The unsupervised Isolation Forest model is trained with normal data, and it showed 98% 

accuracy with the complete test set. Despite not having seen jamming examples during 

training, it achieved a recall of 0.94 and a precision of 0.92, effectively identifying 

abnormal patterns. However, the visual comparison of predictions of IF with the 

positioning data demonstrated stimulating result; the model has captured not only the 

data absence period but also the data degradation intervals. Therefore, the reason 

behind false predictions of this unsupervised model can be a problem with the labelling 

algorithm.  

 
Feature importance analysis from the XGBoost model revealed that G1_min and 

G1_mean were the most influential features, which contribute over 90% of the model’s 

predictive power. This suggests that the G1 frequency band is most sensitive to jamming 

effects and can be prioritized in future models for better performance and efficiency.  

 



62 

 

   

 

With the above result, the supervised learning method, XGBoost, is highly effective in 

controlled environments where explicit ground truth is available. The model performs 

well in real-time jamming detection scenarios where historical patterns of normal and 

jammed signals can be used to train reliable classifiers. The model is suitable for 

applications such as transportation systems where the performance tracking and data 

labelling are feasible.  

 

The unsupervised model, IF, is well-suited for situations where the absence or 

unreliability of ground truth is present. Further, it is applicable for early jamming 

detection purposes. This model can be highly recommended for real-time applications, 

such as autonomous driving, for early warning of jamming. However, the IF model can 

be unreliable in high-noisy environments because it may lead to false predictions.  

 

Authors, Ghanbarzadeh et al., (2025) have carried out a comprehensive analysis of GNSS 

jamming detection and classification using machine learning methods. They achieved an 

accuracy of 98.93% with the ResNet18 model using transfer learning, while the XGBoost 

model developed in this study achieve