FlagThis

Abstract

Federated learning (FL) enables aircraft feet operators to jointly train remaining-useful-life (RUL) models from engine sensor telemetry without sharing raw data. Among the challenges associated with heterogeneous participants, this study focuses on two complementary forms: benign heterogeneity, where honest operators observe diferent operating conditions and fault modes, and adversarial heterogeneity, where a compromised operator submits poisoned updates. We present a controlled, safety-oriented study of how federated training and aggregation algorithms behave under both, using a multi-task one-dimensional convolutional neural network on a structurally non-independent-and-identically-distributed (non-IID) partition of the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) benchmark, and evaluate four benign-heterogeneity remedies together with a fve-attack by four-aggregator matrix that includes a physically motivated, failure-masking sensor-value backdoor. Shared-representation personalization closes about 70% of the local-to-centralized root-mean-square-error gap, versus about 21% for proximal regularization and 10% for server-side reweighting. Crucially, the backdoor reaches 94.9% attack success against standard averaging while clean accuracy stays statistically unchanged: accuracy alone cannot certify a safe model, and attack success must be measured explicitly. Robust aggregation with Krum cuts attack success by an order of magnitude and is the only evaluated aggregator to survive coordinated attackers, whereas personalization alone ofers no protection. Stacking the two yields a composed defense that restores robustness (attack success 2.8%) at only a small accuracy cost, refecting an inherent tension between robust update selection and collaborative representation learning. Findings hold across client counts and a harder six-condition dataset, and all code and data partitions are released for reproducibility.