Problem
A security team monitors an industrial process, such as a water plant, with an autoencoder. The model flags a window as anomalous, but it cannot say which of dozens of sensors and actuators caused it. Comparing input and output element by element fails, because an anomaly in one input can show up in other output positions. Without an explanation the analyst cannot investigate or define where the model is safe to use.
Example from the article: on the SWaT water treatment testbed, the team had to explain each of 36 known attacks by naming the tank level or flow signals behind it.
Solution in brief
Treat explanation as a separate step after detection. For continuous signals, freeze the trained autoencoder and search for the smallest sparse correction to the input that removes the reconstruction error; the changed signals are the culprits. For discrete signals use a restricted Boltzmann machine, and for any model use the model-agnostic LIME and SHAP.
Questions this case answers
- How to find which input signals caused an anomaly flagged by a reconstruction-based detector when the signals mix continuous and discrete values
- How to test an anomaly explanation method when real attack data is unavailable
- How to find which signals caused an anomaly flagged by an autoencoder
- How to explain anomalies in industrial telemetry at the feature level
Tools and methods
- Autoencoder — detects group anomalies where each point looks normal but the combination does not.
- Sparse input optimization (Ikeda et al.) — gradient search over the input with an L1 penalty, to find the signals that drive the error.
- Restricted Boltzmann machine (RBM) — memorizes normal combinations of binary states and flags unusual ones.
- LIME and SHAP — local, model-agnostic feature contributions for each anomaly.
- SWaT testbed — public water treatment dataset with labeled attacks.
Google or a one-time email code. No password.