Autonomous vehicles can be imperfect — As long as they’re resilient

(Press-News.org) Researchers from three of Virginia’s premier universities, including the University of Virginia’s Homa Alemzadeh, aim to take the risk out of self-driving vehicles by overcoming inevitable computer failures with good engineering.

The trio will share a $926,737 National Science Foundation award to identify when and where autonomous vehicle systems are most vulnerable to safety-critical failures. They plan to use this knowledge to design ways to efficiently mitigate potential safety hazards and enhance the overall system resilience. 

Alemzadeh, an associate professor of electrical and computer engineering in UVA’s School of Engineering and Applied Science, is collaborating with William & Mary professor of computer science Evgenia Smirni, and George Mason University assistant professor of computer science Lishan Yang, the lead investigator.

Alemzadeh said recent studies show a significant portion of “disengagements,” in which the autonomous system is turned off by the system or a human driver for safety reasons, happen when the machine learning-based AI makes wrong decisions, or doesn’t make them in time.

“We are particularly interested in studying disengagements and safety incidents due to transient hardware faults, temporary loss of network connection or software errors,” she said.

Hunting Elusive Vulnerabilities These often self-correcting events may cause only a momentary disruption, and then they’re gone, making them hard to find and diagnose.

For safety and reliability, self-driving vehicles depend as much on the software “controller” that makes and executes decisions for autonomous operation and machine learning components as physical parts such as sensors and brakes. 

If transient faults are activated at a critical operational time, they can propagate throughout the system’s hardware and software layers, evade existing safety checks and create hazards.

“We aim to look at the end-to-end system — from input to output — to investigate the critical fault locations within the hardware and software, as well as the system contexts that lead to activation of faults and safety hazards,” Alemzadeh said.

The researchers will focus on improving controller and machine learning components to prevent accidents. Using cross-layer reliability analysis, they plan to strategically target the most critical faults in the vast, complex software code and hardware that underlie the controller and machine-learning models.

Test-Driving Solutions in Real Time Based on when and where they find vulnerabilities, the team will design protection mechanisms —such as automatically correcting transient faults or mitigating unsafe vehicle operations by automatically slowing down — that can be applied selectively at different times and locations to ensure safety while maximizing efficiency.

The team will validate their solutions through closed-loop testing in which the autonomous system and its safety features will be tested in real time using driving simulations under varying weather, road and traffic conditions. They will concurrently simulate faults and errors to assess those impacts and the performance of their solutions.

The three-year NSF project, End-to-End Resilience in Autonomous Driving Systems: Strategic Vulnerability Assessment and Mitigation, follows a previous collaboration between Alemzadeh and her colleagues. Toward Trustworthiness in Autonomous Vehicles, funded by the Commonwealth Cyber Initiative Coastal Virginia regional node, laid the groundwork for the current project.

With four regional nodes, the Commonwealth Cyber Initiative is a coalition of state higher education institutions that partners with government, industry and non-government organizations to make Virginia a global leader in cybersecurity through research, innovation and workforce development. 

Alemzadeh also is a member of UVA Engineering’s Link Lab, a multidisciplinary research and education center for cyber-physical systems, and holds a courtesy appointment in computer science.

END