When Stealth Requires Memory: Budgeted Attack Scheduling under a Whiteness Constraint

Abstract

We pose stealthy attack scheduling on a sensor-to-estimator link under a resource constraint with a budget \barΓ on the fraction of corrupted transmissions and a model-free whiteness constraint on the received innovations. For a discrete-time linear plant with non-Gaussian noise, the worst attack, innovation sign flip, preserves the innovation magnitude and is exactly stealthy against every magnitude-measurable detector, the damage-optimal schedule being a memoryless threshold; but the tail concentration that maximizes damage also manufactures serial correlation, where an innovation-whiteness monitor gains power. We dualize the whiteness constraint and show the optimum is a threshold rule on corrected innovation energy using memory of recent magnitudes and the previous decision. We further trace that correction to a second degree of freedom and set the damage by the location of the firing set in magnitude space and set exposure by the boundary density of its run structure. We realize it as a hysteresis set by one split-conformal order statistic without any plant model, using one counter and two comparisons per step. It keeps 80--96\% of the memoryless damage at 2.3 to 71 times lower lag-one whiteness power, validated on a real truck CAN record.