Black-Box Adversarial Attacks on Smart Grid Stability Score Prediction and a Stochastic Quantile Smoothing Defense
Black-Box Adversarial Attacks on Smart Grid Stability Score Prediction and a Stochastic Quantile Smoothing Defense
Papaer Accpeted»
Emad Efatinasab
·
Mirco Rampazzo
.
Alessandro Brighente
·
Chuadhry Mujeeb Ahmed
·
Mauro Conti
Please, cite this work when referring to the paper.
???
Regression-based stability prediction is a core component of decentralized smart grid control, yet its security under adversarial data-injection attacks remains largely unexplored. Unlike prior work that focuses on binary stability classification, many operational systems rely on continuous stability scores that directly influence control decisions. In this paper, we study black-box adversarial attacks against regression-based stability prediction under realistic query and injection budget constraints. We adapt and evaluate multiple query-based derivative-free attacks designed to directly maximize regression output distortion rather than misclassification. We further introduce an uncertainty-aware sampling strategy that allows an attacker to concentrate a limited injection budget on the most vulnerable inputs, significantly amplifying attack impact. To mitigate these attacks, we propose Stochastic Quantile Smoothing (SQS), a lightweight, query-side defense that randomizes model outputs using secret quantile sampling and input smoothing. SQS returns a single stochastic but plausible scalar per query, deliberately obstructing gradient and expectation-based black-box optimization while preserving utility for legitimate users. Extensive experiments on smart grid stability prediction show that SQS substantially reduces the effectiveness of black-box attacks by increasing adversarial query cost, while incurring only modest degradation in predictive performance.