Federated Learning Enhances Aircraft Engine Prognostics, Resisting Attacks.

Researchers have investigated how to make federated learning (FL) more robust and personalized for predicting the remaining useful life (RUL) of aircraft engines. This is crucial for fleet operators who want to collaboratively train models without sharing sensitive raw data, but face challenges from diverse operating conditions (benign heterogeneity) and malicious actors (adversarial heterogeneity). The study used a multi-task convolutional neural network on a non-IID partition of the C-MAPSS benchmark. It found that shared-representation personalization significantly closed the accuracy gap compared to centralized models, while robust aggregation methods like Krum were essential for resisting adversarial attacks, including a physically motivated sensor-value backdoor. Critically, combining personalization with robust aggregation restored strong protection against attacks with minimal accuracy cost, highlighting a trade-off between collaborative learning and update selection.