Federated Learning Boosts Aircraft Engine Prognostics, Enhances Security

A new study investigates the application of federated learning (FL) to predict the remaining useful life (RUL) of aircraft engines, allowing fleet operators to collaboratively train models without sharing sensitive raw data. The research tackles two critical challenges: benign heterogeneity, where different operators have varying engine conditions, and adversarial heterogeneity, involving malicious data poisoning. The findings indicate that a shared-representation personalization approach substantially closes the accuracy gap compared to centralized models. Furthermore, while standard aggregation methods are vulnerable to physically motivated sensor-value backdoor attacks, robust aggregators like Krum can significantly reduce attack success rates. The most effective solution involves combining personalization with robust aggregation, which maintains high accuracy while providing strong protection against coordinated attackers. This approach demonstrates a crucial trade-off between collaborative learning and robust update selection, with results remaining consistent across different client counts and datasets. The study emphasizes that model accuracy alone is insufficient for certifying safety, and explicit evaluation of attack success is necessary.