Researchers analyze the adversarial robustness of state space models (SSMs) for time-series forecasting, proving Spacetime SSMs can represent optimal Kalman predictors for autoregressive processes. They formulate robust design as a Stackelberg game and derive closed-form bounds showing how instability and decoder dimension amplify vulnerability. Experiments show model-free attacks without gradient access can degrade accuracy by at least 33%.
Safety
Adversarial Robustness of Deep State Space Models for Forecasting
State space models for time-series forecasting have a fundamental vulnerability to model-free adversarial attacks that can degrade accuracy by 33%, driven by amplification through instability and decoder dimension.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.LG (Machine Learning)BY sys://pipeline
Tags
safety