Omrilevi123/gnss-guardian

GNSS Guardian: GPS Jamming and Spoofing Detection

A synthetic dataset for detecting GPS attacks (jamming and spoofing) from a phone's satellite signal. Built as the final project for Introduction to Data Science.

The goal is simple: detect when your GPS is being jammed or spoofed, so you can switch to a safe satellite system such as GLONASS instead of trusting a wrong position.

The data generation and EDA notebook is included in this repository.

Background

GPS is used almost everywhere: navigation, ride hailing, shipping, drones, and even time synchronization for banks and power grids. It has two common attacks:

• Jamming: a cheap transmitter floods the GPS frequency with noise, so the phone loses its position.
• Spoofing: an attacker sends fake satellite signals, so the phone reports a wrong position while believing it is fine.

Most phones cannot tell that they are under attack. The goal of this project is a model that reads the phone's raw satellite signal and decides whether it looks Normal, Jammed, or Spoofed, plus some extra context: the attack direction, which satellite systems are affected, similar past cases, and a short written report.

GPS terms in plain language

A few signal terms appear throughout. Explained in data science words:

• Signal strength (called C/N0): how strong each satellite's signal is. Higher is better.
• Gain (called AGC): how much the receiver amplifies what it receives. It goes up when there is interference, which is a strong clue for jamming.
• Satellites in view: how many satellites the phone is tracking at that moment.
• Constellations: the national satellite systems the phone uses at the same time (GPS from the US, GLONASS from Russia, Galileo from Europe, BeiDou from China).
• Azimuth and elevation: the compass direction and the height in the sky of a satellite.

What is in the dataset

• 14,850 rows, 26 columns, 3 balanced classes (Normal, Jamming, Spoofing, 4,950 each).
• Each row is one moment in time for a receiver, summarizing 8 to 12 satellites.
• Each row also has a short text description of its scenario, used later for a recommendation feature.

import pandas as pd
df = pd.read_parquet("gnss_dataset.parquet")

How the data was generated

The data was made in two steps:

1. Scenarios from a language model. We prompted a Hugging Face model (Qwen2.5-1.5B-Instruct) to write 450 different attack scenarios: the attack type, the environment (city, forest, open sky, and so on), how severe it is, the direction, and a short text description.
2. Turning scenarios into signals. A rule based generator turned each scenario into a time series of signal readings using simple physics. For example, satellites high in the sky give stronger signals, jamming lowers signal strength and raises gain, and spoofing adds fake strong signals.

The generation notebook is in this repository.

Features

Group	Columns
Labels and context	attack_type, severity, environment, scenario_id, time_step, description
Signal strength	cn0_mean, cn0_std, cn0_min, cn0_max
Gain	agc_db
Structure	num_satellites, elevation_mean, cn0_elev_corr
Anomaly	position_jump, bearing_confidence
Per satellite system	cn0_gps, cn0_galileo, cn0_glonass, cn0_beidou, const_cn0_spread
Direction	est_source_azimuth, est_source_elevation, true_attack_azimuth, true_attack_elevation

The feature cn0_elev_corr is the correlation between a satellite's height in the sky and its signal strength. In a normal signal these go together (higher satellites are stronger). Spoofing breaks this pattern, so it is a useful clue.

Exploratory Data Analysis

The charts below are also produced in the EDA notebook.

Which features separate the classes. Using mutual information (how much a feature tells you about the label), gain is the strongest single feature at 0.86, followed by the signal strength features at about 0.55 each. The number of satellites scored 0.000, which is a good sign: it means the label was not accidentally leaked into that column.

The hard pair is Normal versus Spoofing. On every single feature, Normal and Spoofing overlap, so no single feature can separate them. This is why a model that combines several features is needed. The signal versus height feature (cn0_elev_corr) is the most useful one for this pair, because spoofed signals do not follow the normal pattern.

Attack direction. The dataset stores an estimated direction of the attack. Spoofing can be located reasonably often (about 48 percent of scenarios within 45 degrees, roughly twice as good as random), because it comes from one transmitter. Jamming cannot be located (about 90 degrees error, which is random), because it is broadband noise coming from everywhere. This matches how the two attacks behave in reality.

Which satellite systems are hit. Some jammers only affect one frequency band. When that happens, GLONASS often survives because it uses a different frequency. The dataset shows this pattern, which supports a simple defense: if only some systems are jammed, switch to a healthy one.

Data quality problems from the language model (a required check). The language model made mistakes while generating scenarios, and part of the EDA was to find and fix them:

• It sometimes left signal values empty.
• It sometimes gave a Normal scenario an attack severity.
• It preferred round numbers for direction (56 percent of directions were multiples of 45 degrees) instead of realistic values.
• It never labeled an attack as low severity.

All of these were identified and corrected during cleaning.

Data hygiene. The final dataset has 0 duplicate rows, missing values only where they are meaningful (a satellite system not in view, or a Normal row that has no attack direction), and all feature values fall in sensible ranges.

Models

• Random forest (for analysis). Reached 99.9 percent accuracy, but only when the data was split by scenario instead of by row. This avoids leakage, because rows from the same scenario are very similar and a random split would let the model cheat. A random split gave the same number, which confirms there is no leakage. Adding noise to the test set dropped accuracy gradually (94.5 percent with mild noise, 84.3 percent with moderate noise), which shows the model degrades smoothly rather than breaking.
• 1D CNN (used in the app). Trained on 6 features in windows of 30 time steps, and tuned to match real phone measurements.
• Recommendation. Scenario descriptions were turned into vectors using three Hugging Face embedding models (MiniLM, BGE, E5). BGE performed best, with 86 percent of nearest neighbors sharing the same attack type. FAISS returns the most similar past cases quickly.
• Report generation. A small Hugging Face language model writes a short incident report: what happened, how severe it is, the likely cause, and a suggested action.

Distribution shift: making it work on a real phone

This was the most important finding. A model trained only on clean synthetic data failed on a real phone. A normal recording from a Samsung Galaxy S23 was labeled as 68 percent attack. Looking at the real data showed why:

Feature	Synthetic (first version)	Real Samsung S23
Signal strength (C/N0)	about 42	about 29
Gain (AGC)	about -85	about -50
Satellites in view	about 10	about 40

The real values sat inside the synthetic attack range, so the model raised false alarms. This is called distribution shift.

The fix had three parts:

1. Retrain the model using realistic ranges taken from the real recording, and vary the device baseline during training so the model does not depend on exact values.
2. Add calibration in the app. The first 30 rows of an uploaded file are used as that device's baseline, and the rest is adjusted to match.
3. Redefine one unstable feature so it measures sudden changes instead of normal satellite motion.

After these fixes, the real recording was labeled correctly. The takeaway is that high accuracy on synthetic data does not prove real world performance. Matching the real data distribution and calibrating per device is what makes it work.

Limitations

• This is a proof of concept based on realistic signal behavior, not a finished product.
• No real attack recordings were used, because transmitting jamming or spoofing is illegal. Attacks are simulated.
• The direction is an estimate. Real localization needs special hardware or a moving receiver.
• Real performance depends on calibrating to each device.

Files

• gnss_dataset.parquet: the dataset (14,850 rows, 26 columns)
• generation_notebook.ipynb: data generation
• eda_notebook.ipynb: EDA and model training
• README.md: this file

Authors

Final project for Introduction to Data Science.