Federated learning improves gps spoofing detection in drone fleets
Receiver-Domain Behavioral Probing for Backdoor-Resilient Federated GPS Spoofing Detection in UAV Networks
Cryptography and Security
Summary
GPS spoofing tricks drones into thinking they are somewhere they are not, which can be dangerous. The authors show that a common method drones use to share learning about spoofing can be tricked by dishonest drones pretending their models are better than they are. They propose a new way to check each drone’s model more carefully based on how it behaves on test data, rather than trusting what it claims. This new method can spot cheating drones and keep the shared detection working well even if some drones lie.
What this means in practice
- •For drone fleet operators: Deploy stronger GPS spoofing detectors that resist dishonest drone inputs by using behavioral checks on learned models.
- •For network security teams: Improve trustworthiness of federated learning systems by verifying client models through counterfactual behavior rather than self-reported accuracy.
Tested on one dataset.
Authors
Will Jedrzejczak, Cole Walther, Dilpreet Gill, Moh Khalid Hasan
Abstract
Federated learning lets a UAV fleet train a shared GPS spoofing detector without raw receiver data leaving any aircraft, and several recent UAV-FL designs weight each client by the validation accuracy it reports about itself. We show this self-report is an exploitable attack lever: two compromised clients of ten that poison part of their data, scale their updates, and inflate their reported accuracy raise backdoor lift to +0.3036, higher than the same attack achieves without lying. We propose receiver-domain behavioral probing, in which the coordinator evaluates every submitted model on counterfactual spoofed samples built by driving each discriminative GPS feature to a benign value, weighting clients by what their models do rather than what they claim. Under independent and identically distributed clients this reduces attacker-induced lift to -0.0265, statistically indistinguishable from an honest fleet, while flagging the compromised aircraft. Unlike Byzantine-robust aggregation it needs no exact attacker count, only an honest majority: when the true count exceeds the configured value, Multi-Krum degrades from +0.0061 to +0.2837 while ours stays near baseline. Evaluation uses one public single-receiver dataset partitioned into simulated clients; we also report where the mechanism fails, under strong client heterogeneity.