MIT researchers boost aircraft engine AI security with federated learning
New federated learning model thwarts 94.9% of adversarial attacks on aircraft engines...
Federated learning lets aircraft fleet operators train remaining-useful-life models from engine sensor telemetry without sharing raw data. A new safety-oriented study tackles two threats: honest clients with different operating conditions and fault modes, and compromised clients sending poisoned updates. Using a multi-task 1D convolutional neural network on a structurally non-IID C-MAPSS partition, researchers compared four remedies for benign heterogeneity and five attacks against four aggregation methods. Shared-representation personalization closed about 70% of the local-to-centralized root-mean-square-error gap, far ahead of proximal regularization (21%) and server-side reweighting (10%). A physically motivated sensor-value backdoor achieved 94.9% attack success against standard averaging while leaving clean accuracy statistically unchanged — showing accuracy alone can't certify safety. Krum reduced attack success by an order of magnitude and was the only evaluated aggregator to withstand coordinated attackers; personalization alone provided no protection. Combining personalization with robust aggregation restored robustness (2.8% attack success) with only a small accuracy cost, revealing a trade-off between robust update selection and collaborative representation learning. Results held across client counts and a harder six-condition dataset, and code and data partitions are released for reproducibility.
- New federated learning framework combines personalization (closing 70% of accuracy gap) with Krum aggregation to reduce adversarial attack success from 94.9% to 2.8%
- Tested on NASA's C-MAPSS benchmark using a multi-task 1D CNN architecture across both benign and adversarial scenarios
- Code and data partitions released for reproducibility under arXiv:2608.04045
Why It Matters
This breakthrough enables secure, collaborative AI for aircraft maintenance while preventing malicious actors from compromising safety-critical predictive systems.