Self-driving security insights
- Self-driving vehicles are becoming more popular on the streets, emphasizing the need for self-driving security to protect vehicles from cyberattacks.
- Virtual scenario-based training is a more cost-effective way of testing efforts, instead of using physical vehicles on the road.
Self-driving technology can save lives, but adding autonomous vehicles multiplies the number of cyberattack targets on the road — endangering not only those in the affected vehicle, but those in nearby vehicles and buildings. Road testing to improve self-driving security is expensive and time-consuming, but Commonwealth Cyber Initiative (CCI) researchers developed a virtual option: a vast digital library of real-world driving scenarios to help manufacturers craft more secure autonomous vehicles.
A team from the Virginia Tech-affiliated company Global Center for Automotive Performance Simulation (GCAPS) is working with researchers at the Virginia Tech Transportation Institute (VTTI) to validate the security and performance of automated technologies virtually. Their project is supported by the Innovation: Ideation to Commercialization program from the Commonwealth Cyber Initiative in Southwest Virginia, which provides funds to commercialize technology generated from research at the intersection of data, security and autonomy.
Virtual, scenario-based testing is a safe, cost-effective alternative to the traditional distance-based validation of autonomous technology, which requires the onerous and sometimes impossible task of driving billions of miles on public roads.
“The library establishes ground truth based on real-world data for vehicle interactions with the road, the infrastructure, vulnerable road users and other vehicles,” said Miguel Perez, associate professor of biomedical engineering and mechanics who leads the VTTI data engineering program.
Reams of real data
To construct a highly reliable virtual training library, the research team is leveraging a unique resource: the massive treasure trove of real-world recorded events in VTTI’s naturalistic driving databases.
For more than 30 years, VTTI has been collecting data from cameras and sensors unobtrusively installed (with permission) in regular cars belonging to regular people all over the country. The data — which includes basic vehicle dynamics, video and sensor information — represents over 70 million miles of real-world driving behaviors exhibited during normal daily commutes, culminating in over 10 petabytes so far.
Over the past five years, GCAPS has worked with VTTI to apply the data to automated driving technology. The team has been developing an algorithm to classify a vehicle’s lateral and longitudinal micro movements by assigning objective values to the displacement of a vehicle at certain rates and distances. Once an event has been classified, it can then be recreated: converted into a simulation-ready dataset that includes trajectories, positions, road features and terrain.
The researchers have been exploring questions like how close do people really drive to each other? Does a person’s age affect their reaction? Would a person in a different region react differently?
In the event of a cyberattack
“It’s a really powerful tool that’s only getting stronger and richer,” said Jonathan Darab, operations director at GCAPS. “Now, with CCI support, we’re utilizing it to enhance the cybersecurity of automated driving technology.”
With this focus, the question set has expanded to include what might happen with vehicle performance in the event of a cyberattack? How will a vehicle compensate for a spoofed sensor? Will the autonomous vehicle be safe on the road if it’s actively being hacked?
“Because the interactions in the events represent real-world data, the accuracy and timing of the interactions provide a highly valid foundation,” said Perez. “As such, the library enables the determination of cybersecurity risk severity associated with vehicles, but in a safe and repeatable way through simulation.”
This capability will help engineers determine the requirements for self-driving security products on automated and electric vehicles, such as the required response lag times to minimize vehicle issues, or the percentage of computing resources that can be safely reassigned to defend a hacked network.
Original content can be found at Virginia Tech.