Adversarial Evasion Attacks on SVM-Based GPS Spoofing Detection Systems.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Robotics

Background:

• Global Positioning System (GPS) spoofing poses a significant threat to autonomous vehicles.
• Machine learning models, such as Support Vector Machines (SVMs), are commonly used for GPS spoofing detection.
• The robustness of these detection systems against sophisticated adversarial attacks is not well understood.

Purpose of the Study:

• To investigate the vulnerability of SVM-based GPS spoofing detection models to intelligent adversaries.
• To develop novel evasion strategies for adversarial GPS signals.
• To assess the effectiveness of these strategies in bypassing SVM detectors.

Main Methods:

• Analysis of the decision boundary of an SVM-based GPS spoofing detection model.
• Development of a data location shift attack and a similarity-based noise attack.
• Simulation of evasion strategies in the CARLA autonomous driving environment.

Main Results:

• A modest positional shift attack reduced detection accuracy from 99.9% to 20.4%.
• Similarity-based noise attacks largely evaded detection while degrading performance.
• A nonlinear cancellation effect was observed between similarity and shift, indicating a detectability-impact trade-off.

Conclusions:

• SVM-based GPS spoofing detection systems are vulnerable to novel evasion attacks.
• Existing defenses may not be robust against intelligent adversaries.
• Enhanced adversarial robustness is crucial for machine learning-based spoofing detection in vehicular systems.