Anomaly Detection using IoT Sensor-Assisted ConvLSTM Models for Connected Vehicles

The safety and security of connected autonomous vehicles' (CAVs) passengers are crucial for the autonomous vehicle industry. A zero-loss and accident-free strategy is leading the way, not only considering a luxurious development and design of the automotive industry but also avoiding the cyber attacks against the vehicle intrusion detection system (IDS) that may be at its best performance keeping attackers from maliciously altering or corrupting the flow of data within the vehicle's internal communication bus. In the Vehicle-To-Everything (V2X) age, there is no guarantee,, that faulty data streams out of critical Electronic Control Units (ECUs) can be kept from leading the autonomous vehicle astray without external help from different sets of IoT sensors in pedestrian-held devices, passenger-held devices, or side-road infrastructure. Therefore, in this research, ConvLSTM deep learning models are proposed to recognize corrupted data streams bundled together with the autonomous vehicle navigation and powertrain data. The efficiency and generalization of the supervised ConvLSTM deep learning models are tested on anomalies within a stack of different in-vehicle sensors assisted by IoT navigation sensor data (e.g. Global Navigation Satellite System (GNSS) location, motion, and orientation data). The proposed models are evaluated using a dataset that consists of many IoT sensors that, although, was not collected through an IoT network infrastructure, being collected through real-world vehicle ride can very well represent a benchmark for any V2X IoT streaming bundle that is used for autonomous vehicle positioning applications in the future. In this paper, we achieved anomaly detection with 0.98 F1-score using different ConvLSTM model designs which is much higher than most state-of-the-art approaches and on par with state-of-the-art deep learning LSTM models.

Automated summary

This study utilizes ConvLSTM deep learning models to identify corrupted data streams in autonomous vehicle data. By using a variety of sensors and IoT navigation sensor data, the efficiency and generalization of the model are improved.