Model-Free Budgeted Attack Scheduling for Cyber-Physical Systems

Qazi Mairaj ud din, Sidra Ghayour Bhatti, Qadeer Ahmed

This letter studies the budgeted scheduling of stealthy false data-injection (FDI) attacks against state estimators in cyber-physical systems. Existing event-based attack schedulers require full knowledge of the plant model and assume the residual distribution is exactly Gaussian -- assumptions that fail for real-world CPS sensor streams whose residuals are heavy-tailed and whose dynamics are unknown to the adversary. We propose a model-free attack-scheduler that replaces the parametric Gaussian threshold with the empirical quantile of a learned sequence autoencoder residual, calibrated from measurements alone without any plant matrices. We prove that the realized attack rate converges almost surely to the target budget under stationary ergodic residuals. Experiments on two synthetic systems and a real heavy-duty truck dataset show that the proposed scheduler tracks the budget to within 1-2% while also preserving the residual magnitude, guaranteeing stealthiness against any residual-based detector. Comparing with the model-based baseline -- granted the true plant and innovation covariance -- mis-realizes the budget by up to 8.96% under heavy-tailed residual distribution, causing the attacker to achieve only 1.37x system degradation when 1.84x is intended.