ARAD: Automated and Real-Time Anomaly Detection in Sensors of Autonomous Vehicles Through a Lightweight Supervised Learning Approach

In this paper, an automated and real-time anomaly detection approach for sensors of autonomous vehicles called ARAD is presented. Automated vehicles gather environmental information through their diverse built-in sensors thus the correctness of this data affects the system’s reliability, directly. Accordingly, anomaly detection schemes are employed to guarantee the correctness of the sensors’ data. Moreover, due to the necessity of real-time operation in automated vehicles, the response time of the anomaly detection unit is important along with its precision. To this aim, in our proposed ARAD a lightweight and hierarchical architecture to detect and classify the anomalies based on their types is employed. Moreover, to enhance the detection capability, ARAD utilizes the data diversity property based on the sequence prediction scheme. After anomaly detection, ARAD mitigates and removes them from the system’s input by its rule-based engine. To meet the precision and real-time requirements of the anomaly detection unit in autonomous vehicles, ARAD has a lightweight sequence prediction structure based on statistical and data-driven methods. To evaluate the effectiveness of our proposed ARAD, several experiments are performed and a performance measurement metric called FoC is proposed to study the contradicted effects of precision and real-time operation in terms of computation overhead, simultaneously. Based on these experiments, ARAD is capable of detecting anomalies efficiently with precision and recall of
$84.6~\%$
and 87%, respectively in real-time while applying low overhead to the system. It also shows 75.6% improvement in terms of computation cost over related methods.