Background
In recent projects, SwRI has demonstrated the ability to remotely exploit automated vehicle sensors. Customers are interested in the vulnerability exploitation but are more focused on solutions to mitigate these threats. The proposed research will focus on detecting erroneous data in real-time on vehicle sensor networks. Cyber-attacks such as remote spoofing and lower tech issues like being obstructed or damaged will be detected using machine learning-based algorithms tailored for anomaly detection. Successful implementation improves security and increase safety in automated vehicles and Automated Driver Assistance Systems (ADAS).
Approach
The approach in this research utilizes GPS, wheel speed, and LIDAR data to develop an anomaly detection algorithm. Three approaches were developed during this research:
Time-magnitude Based Thresholding
Temporal Autoencoder (TAE)
Density-Based Spatial Clustering of Applications with Noise (DBSCAN)
Various faults were simulated by modifying sensor data to test the performance of each approach. The vehicle used for the project is shown in Figure 1.
Figure 1: SwRI-Owned Lincoln MKZ Outfitted with LIDAR, GPS, Wheelspeed, and other sensors.
Accomplishments
Of the three approaches explored, the time-magnitude approach yielded the best results. The thresholding approach was able to detect most of the simulated faults and it was also able to overcome deficiencies in the data set such as noise. The results achieved during this research shows a promising aptitude for anomaly detection on direct sensor data in autonomous vehicle platforms. With further improvements, this solution incorporate a larger variety of data and sensors. The results of this research leads to a solution that addresses concerns of cyber-security and information assurance in autonomous vehicles.