Protecting autonomous systems from GPS spoofing with a machine learning-driven approach

Abstract

With the rapid evolution of interactive multimedia systems, ensuring strong security measures has become increasingly vital. Autonomous platforms, such as drones, are vulnerable to sophisticated cyber threats, including jamming and spoofing attacks. One common spoofing strategy involves manipulating Global Positioning System (GPS) signals. By broadcasting counterfeit signals, attackers can deceive drone navigation systems. To mitigate these risks, this research introduces a machine learning-based framework aimed at intelligently detecting spoofing attempts. The approach employs signal characteristics that reflect variations in jitter, shimmer, and frequency modulations, rooted in mathematical analysis. A private dataset is used for this work. This dataset, developed by our research team, is collected under varied environmental conditions, such as during daylight and in low-light settings, over multiple sessions. It categorizes signal statistics into three distinct ranges: the initial and final segments suggest spoofed signals, whereas the middle range corresponds to genuine ones. Further, using signal reception strength (SRS) values, additional data was sourced from trusted Long Range Wide Area Network (LoRaWAN) devices. This information supports the development of a singular-class support vector machine (S-CSVM) classifier for spoofing detection. For training purposes, the dataset was partitioned into two subsets: a training set (Ttrain[jls-end-space/]) and a testing set (Ttest[jls-end-space/]). The performance of the model is evaluated using standard metrics, including precision, recall, F-score, and overall accuracy. By utilizing all available features, the model achieves its highest scores: 99.99% precision, 99.77% recall, and 99.95% F-score. The highest accuracy of 99.22% is achieved when all features are selected, and the distance between LoRaWAN devices and the monitoring device ranges from 6 to 8 m, outperforming the results obtained through feature selection in the ablation study. A thorough evaluation further highlights how the proposed ML-based solution outperforms existing methods.

Keywords: Cyber threats; GPS, Machine learning, Mathematical modeling, Security, Spoofing, UAVs