Researchers in China have developed a lightweight module to secure global navigation satellite system (GNSS) signals used in existing UAV autopilots, writes Nick Flaherty.

Civilian UAVs rely on unencrypted GNSS navigation, making them easy prey for spoofers who broadcast slightly stronger fake satellite signals. Once a UAV locks onto the counterfeit constellation, it miscalculates its position and veers off mission. Existing countermeasures demand expensive multi-frequency receivers or continuous links to cellular base stations to provide secure reference data – requirements that are impractical for low-cost agricultural or delivery platforms.

The Motion-State-Series Trajectory Prediction and Online Anomaly Detection (MSSTP-OAD) system, developed by researchers from Beijing University of Posts and Telecommunications and Pengcheng Laboratory to fit into the ArduPilot autopilot, reframes the problem as short-term trajectory forecasting.

At the start of the process, a type of AI called a stacked Long Short-Term Memory (LSTM) framework is trained on flight logs – straight segments, turns, climbs and loiters – recorded in the open-source software-in-the-loop simulator. Each training sample is a 20-step sequence (5 s at 4 Hz) of motion-state vectors that fuse position, velocity, acceleration, attitude and magnetic field readings, and the network learns to predict the next five positions from the past 20 states.

During flight, the algorithm works in two stages. The first is rapid screening, where every small time slot constructs a lightweight motion vector and feeds it to a model, called an ensemble model (E1), for a quick anomaly assessment. At the end of the detection window, a high-dimensional vector that adds LSTM-predicted positions is sent to a second ensemble model (E2) that combines various sources of data under a strict majority-vote rule, sharply reducing false positives.

Tests on 30,000 flight segments (half normal, half with 10–100 m horizontal spoofing offsets) showed that the trajectory prediction varies between 0.996 for benign operation and 0.994 when under attack. The root mean square error remains at under 5 m, even when under attack.

After an alarm, a simple ‘return-to-waypoint’ manoeuvre flew 26% less extra distance than the baseline method.

The current findings are based on the software-in-the-loop simulations, and the algorithm is being tested with software-defined-radio spoofers to quantify robustness against real-world multipath, atmospheric delay and receiver clock drift.

The team aims to fuse magnetometer and barometer data to counter potential IMU spoofing, and apply quantisation-aware training to minimise the weights used in the LSTM framework to reduce the memory and firmware overhead because the end goal is a drop-in firmware patch for the PX4 hardware and ArduPilot autopilot. This would allow retrofitting of UAVs with a low-cost, zero-extra-hardware shield against GPS manipulation.