With the recent rapid growth in autonomous vehicles (AVs), researchers are also paying more attention to the security risks involved. As systematic approaches have been proposed for crafting adversarial attacks against the artificial intelligence (AI) perception systems that power AVs, it is imperative to develop effective defences. Most current countermeasures assume that the attackers are ignorant of the defence mechanisms, but this allows attackers to easily overcome the defence once they obtain information on the defence. In a broader sense, the conventional paradigm of designing defences based on pre-defined attack characteristics cannot address adaptive attackers who advance their attacks.

To address the challenge, our project centers its research around a different defence paradigm called the moving target defence (MTD)* approach. The proposed MTD approach changes the defence’s deep neural networks more rapidly than the attacker’s process of understanding the defence. By doing so, attacks that were designed for a particular defence become obsolete and less effective, the exposure of vulnerabilities and opportunities for attacks become limited, and the complexity and cost for attackers are increased, making it more difficult for attacks to compromise the AI perception systems.

*Recognised by the U.S. Federal Networking and Information Technology Research and Development (NITRD) Program as a “cybersecurity game-change R&D recommendation.”