Adaptive Ensemble Methods for Tampering Detection in Automotive Aftertreatment Systems
Control and diagnostic processes in modern vehicles incorporate nowadays a wide set of functionalities to preserve the vehicle’s health. Automotive vehicles contain embedded systems that must perform a diverse palette of tasks, ranging from less critical tasks (e.g., audio/video media control), to crucial ones, such as controlling the engine, fuel consumption, or the aftertreatment system. This paper identifies and addresses one emerging threat, namely, automotive tampering. Tampering denotes a procedure that changes the behavior of the system to gain financial or functional advantages, without damaging the system and without triggering the built-in safety features of the vehicle. Numerous studies show a growing number of tampered vehicles worldwide and considering that tampered vehicles contribute to air and atmosphere pollution, tampering remains a serious environmental threat. This paper proposes two ensemble-based approaches for tampering detection, both using Long Short-Term Memory neural network predictors, together with Cumulative Sum and Histogram distance-based detectors. Additionally, an Adaptive Majority Weighted Voting fusion methodology is proposed, that considers the historical decisions of the detectors. Experimental results are based on three unique datasets that incorporate a multitude of tampering scenarios. The results prove the efficiency of the proposed ensembles, with a 0% false alert rate and up to 100% detection rate, even when dealing with intelligent tamperers, and even in comparison with state-of-the-art tampering detection solutions. Moreover, this paper offers resource consumption and scalability measurements on a reference embedded system, further demonstrating the integrability of the proposed techniques in a real embedded environment.