Artificial Intelligence To Protect Drones From GPS Spoofing Attacks

Drones have rapidly become essential to modern industries, supporting logistics, precision agriculture, environmental monitoring, and emergency response. Yet their dependence on satellite navigation exposes them to one of the most serious digital threats: GPS spoofing. This attack deceives the drone’s navigation system by sending counterfeit GPS signals, forcing it to change course or lose control. Researchers from King Abdulaziz University and Yeungnam University have developed an artificial intelligence–based solution that detects such attacks in real time.

The proposed system integrates MobileNet, a lightweight deep learning model, with a statistical method known as Chi-square feature selection, allowing small drones to identify fake signals with over 98% accuracy. Unlike conventional hardware-intensive protection, this software operates directly on the drone’s processor, analyzing features such as signal strength, Doppler shift, and timing deviations. By recognizing anomalies within milliseconds, it ensures stable navigation without additional sensors or hardware. The system provides a scalable and cost-effective solution for commercial, civilian, and defense drones, offering a crucial safeguard against electronic interference and cyber manipulation in increasingly autonomous air operations.

This AI-driven security system is designed to detect GPS spoofing by combining deep learning with intelligent signal analysis. It is based on MobileNet, a compact convolutional neural network architecture optimized for low-power and embedded devices. The researchers trained the model on a large dataset containing both authentic and spoofed GPS signals collected under different conditions, including rural, urban, and high-interference environments. Before training, the Chi-square feature selection algorithm was applied to extract the most statistically significant parameters influencing signal authenticity.

From thirteen measured variables, eight—such as signal-to-noise ratio, pseudorange, Doppler frequency, and pseudorange increment—were chosen for model input.