Security and Safety in Environmental Perception Systems with Applications to Autonomous Systems

提出的研究计划旨在提高自主系统应用环境感知系统的安全性。增强自动驾驶汽车（AVs）等应用中使用的环境感知系统，将有助于塑造智能、安全的自动和半自动交通的未来。车辆碰撞每年在全世界造成5000多万人受伤，100万人死亡，而且往往是人为失误造成的。此外，由于视力障碍和其他残疾，发达国家20%以上的成年人无法驾驶。有鉴于此，自动驾驶汽车为汽车行业提供了一种新的移动媒介。在过去的十年里，自动驾驶系统的发展取得了重大进展。商用车提供了辅助和自动停车、自适应巡航控制、甚至公路驾驶等技术，实现了高度的自主。然而，这需要安全可靠的环境感知系统，能够在全天候条件下（如雾、雨、雪）运行，并准备好应对任何网络和对抗性攻击。为了实现这一目标，可以在车辆上安装多个不同的传感器，如雷达、摄像头和激光雷达，以感知车辆周围的环境，以提高可靠性和安全性。自动驾驶汽车驾驶决策的可靠性依赖于感知数据，而感知数据是潜在的威胁面。开发一个框架来保护AV系统免受这些攻击是任何AV系统不可缺少的组成部分。我们将开发和设计基于传感器数据完整性验证的技术，以增强传感器数据的安全性。传感器数据验证可以使用加密或基于数据隐藏的技术（如水印技术）来解决。此外，在过去的几年里，人工智能（AI）的采用在很大程度上改变了自动驾驶汽车感知系统的能力。众所周知，人工智能组件极易受到各种攻击，这些攻击可能会损害自主系统的正常功能。对抗性攻击在输入中产生微小的扰动，这些扰动可能不会被人类检测到，但可能导致人工智能模型以高置信度进行错误分类。在本研究项目中，我们打算解决自动驾驶汽车数字系统中与人工智能组件相关的潜在风险，包括环境感知和路径规划，以增强自主感知系统的安全和保障能力。这将通过采用多模态传感器，结合先进的信号处理和人工智能技术来实现。这一研究将对自动驾驶汽车系统的可靠性和安全性产生影响。随着时间的推移，这项工作将影响自动驾驶汽车在加拿大的广泛部署。

The proposed research program is aimed at increasing safety and security in environmental perception systems of applications in autonomous systems. Enhancing the environmental perception system used in applications such as autonomous vehicles (AVs) will help to shape the future of smart and safe autonomous and semi-autonomous transportation. Vehicle crashes cause over 50 million injuries and 1 million deaths per year worldwide, and are often attributable to human error. In addition, more than 20% of adults in developed countries are unable to drive, due to visual impairment and other disabilities. In light of this, AVs offer a new mobility medium in the automotive industry. Over the past decade, significant progress has been made in the development of autonomous driving systems. Commercial vehicles have offered technologies such as assisted and automated parking, adaptive cruise control, and even highway piloting, achieving a high level of autonomy. However, this necessitates secure and reliable environmental perception systems capable of operating in all-weather conditions (e.g., fog, rain, snow), and ready to deal with any cyber and adversarial attacks. To achieve this, multiple different sensors such as radar, camera, and LiDAR may be installed on the vehicles to sense the surrounding environment of a vehicle to improve reliability and safety. The reliability of AVs' driving decisions is dependent on sensory data which is a potential threat surface. Developing a framework to protect AV systems against these attacks is an indispensable component of any AV systems. We will develop and design sensor data integrity verification-based techniques to enhance the security of sensor data. Sensor data verification can be addressed using encryption or data hiding based techniques such as watermarking techniques. In addition, in the past few years, adopting artificial intelligence (AI) has transformed the capability of perception systems in AVs to a great extent. AI components are known to be highly vulnerable to a variety of attacks which could compromise proper functionality of autonomous systems. Adversarial attacks make small perturbations in the input which may not be detected by humans yet may lead to a misclassification with high confidence by AI models. In this research program, we intend to address the potential risks related to AI components of digital systems in AVs, including environment perception and path planning, in order to enhance the safety and security capabilities of autonomous perception systems. This will be achieved by employing multi-modal sensors in conjunction with advanced signal processing and AI techniques. This research will lead to impact on reliability and security of AVs systems. Over time, this work will influence widespread deployment of AVs across Canada.