← Back to Machine Learning (Statistics)
stat.ML

When does Shapley value help detect sensor faults?

Xubin Fang, Rick S. Blum

May 30, 2026

Shapley values have become popular for identifying which sensors are faulty in systems. This paper rigorously analyzes when this approach actually works: for independent sensors, it's mathematically equivalent to a much simpler test, but for correlated sensors (like paired measurements), it can be either better or worse depending on correlation strength and attack type. Combining both methods yields optimal performance in some cases.
Published as Statistical Analysis of using the Shapley Value for Sensor Anomaly Localization with Accurate Classifiers arXiv:2606.00867
Read the original paper →