Safe Journeys | Scalable Algorithms and Tools for Safety and Security Assurance of Autonomous Vehicles
Akshay Dhonthi Ramesh Babu is a PhD student in the department Formal Methods and Tools. (Co)Promotors are prof.dr. M. Huisman and dr.ing. E.M. Hahn from the faculty of Electrical Engineering, Mathematics and Computer Science and dr.ing. V. Hashemi from AUDI AG.
The ever-increasing complexity of road transportation has driven the need for continuous technological advancements in the automotive industry. One major innovation is autonomous driving vehicles, which improve safety, efficiency, and mobility and are changing the future of transportation. The overarching goal of these advancements is to reduce human error and prevent accidents, ultimately enhancing road safety.
As we progress toward fully autonomous vehicles, it is critical to ensure the flawless operation of autonomous driving systems in safety-critical scenarios to prevent accidents. Achieving this requires a focus on both safety and security. Safety involves safeguarding the vehicle and its surroundings: protecting humans, infrastructure, and other elements from potential harm. It also encompasses ensuring compliance with traffic laws, safety standards, and regulatory requirements. Security, conversely, pertains to safeguarding the vehicle from external threats or malicious attacks that could compromise its functionality. This dissertation seeks to address the following fundamental question:
How can we ensure the safety and security of learning-based autonomous driving systems in complex, safety-critical environments?
At the core of autonomous driving is a continuous, real-time loop of sensing, planning, and acting. Sensing refers to the vehicle's ability to accurately perceive its environment and interpret the behavior of other road users. Planning involves generating a safe and optimal driving path based on the sensed information, ensuring that the vehicle poses no risk to other road users or the surrounding environment. Acting is the process of executing these plans through driving commands that guide the vehicle along the intended path. The key challenge for ensuring their reliable performance across all driving scenarios lies in validating these components.
This thesis aims to develop robust testing and validation methodologies for sensing and planning systems. Note that acting systems are out of the scope of this thesis. In the context of sensing systems, we address the security vulnerabilities in deep learning models. These systems may harbor hidden \emph{backdoors} that can compromise their performance. In the context of planning systems, we present our approaches for addressing the safety challenges posed by path planning systems. These systems, which learn from the data, generate routes to achieve desired goals using multiple trajectories obtained from human demonstrations or previous driving data. However, the generated paths often fail to meet safety standards, as they may not adhere to road rules, consider real-time obstacles, or satisfy safety requirements.
This dissertation addresses the safety challenges associated with autonomous driving technologies and presents innovative solutions to overcome these issues. The proposed solutions, specifically backdoor identification, backdoor mitigation, and path optimization, represent critical advancements that bring us closer to achieving the safe deployment of fully autonomous vehicles.
