EAGER: SaTC: Privacy-Preserving Convolutional Neural Network for Cooperative Perception in Vehicular Edge Systems

协同感知使车辆能够相互交换传感器数据，实现协同目标检测和分类，这对提高自动驾驶汽车的安全性非常有用，因为它们经常受到盲点和/或闭塞的影响。然而，如果没有适当的数据保护措施，由于担心潜在的隐私泄露，很少有车辆可能会共享传感器数据。车辆可以在与其他车辆共享传感器数据之前对其进行加密，但数据加密仅在传输过程中保护数据安全，即接收车辆在解密数据后仍然可以访问隐私和敏感信息。为了解决这个问题，设计了一个保护隐私的卷积神经网络（CNN），以便融合和处理车辆生成的加密传感器数据，以产生基于密文的目标检测结果。接收车辆虽然从接收到的密文中获得有意义的目标检测结果，但无法恢复原始数据，从而保护了与其他车辆共享数据的隐私性。所提出的技术是一种通用的解决方案，可以扩展到为其他机器学习方法提供数据隐私保护。这个项目将是革命性的，因为它将使车辆之间安全地共享有用的传感器数据。这对扩大自动驾驶车辆的视野和视野，保障公共安全，推进智慧交通和国家繁荣，都是非常有益的。该项目提供了广泛的研究活动，包括CNN分析，安全算法设计和硬件编程。研究人员将积极地让不同背景的学生参与这个项目。将作出特别努力，增加代表性不足的学生研究人员的参与。每年将举办“互联和自动驾驶汽车研讨会”，将该项目的研究成果推广给产业界和研究界，造福社会。该项目的目标是了解在实现自动驾驶汽车之间的协同感知时面临的安全和隐私挑战，然后设计一个有效且高效的隐私保护CNN来处理来自多辆汽车共享的加密传感器数据。利用加性秘密共享技术，首先将传感器数据随机分割并加密为两个密文，然后由两个非串通的边缘服务器以加密格式进行处理。为了处理加密数据，边缘服务器使用了一种保护隐私的CNN模型，该模型安全地实现了原始CNN中的所有层，包括安全卷积层、安全激活层、安全池化层、安全区域提议网络层和安全全连接层。由于边缘服务器仅拥有传感器数据的部分视图，因此数据中包含的敏感信息受到保护。为了优化保护隐私的CNN模型，设计了跨层优化和基于现场可编程门阵列（FPGA）的加速技术，以提高目标检测和分类的速度，同时保持低能耗。所提出的系统将作为使用CNN保护数据隐私的令人信服的概念证明，从而为广泛采用保护隐私的CNN处理传感器数据打开了大门。源代码和保护隐私的CNN模型，以及它的训练和测试数据集，将公开提供，作为推动车辆边缘系统数据隐私保护创新研究的催化剂。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Cooperative perception enables vehicles to exchange sensor data among each other to achieve collaborative object detection and classification, which is extremely useful to enhance autonomous vehicles' safety as they frequently suffer from blind spots and/or occlusions. Without proper measures in place for data protection, however, very few vehicles are likely to share their sensor data, due to the concerns on potential privacy leakage. A vehicle can encrypt its sensor data before sharing it with others, however, data encryption only protects data security during transmission, i.e., a receiving vehicle can still access private and sensitive information after it decrypts the data. To address this issue, a privacy-preserving convolutional neural network (CNN) is designed so that encrypted sensor data generated by vehicles can be fused and processed to produce ciphertext-based object detection results. Although a receiving vehicle obtains meaningful object detection results from received ciphertexts, it has no means to recover the original data, thus protecting the privacy of data shared from other vehicles. The proposed technique is a generic solution that can be extended to offer data privacy protection for other machine learning methods. This project will be transformative because it will enable vehicles to securely share useful sensor data between each other. Such an advance can be extremely beneficial for extending the line of sight and field of view of autonomous vehicles, which secures the public safety and advances smart transportation and national prosperity. This project offers a wide variety of research activities, ranging from CNN analysis, security algorithm design, and hardware programming. The investigators will actively engage students with various backgrounds in this project. Special efforts will be made to increase the participation of underrepresented student researchers. A workshop on Connected and Autonomous Vehicles will be organized annually to promote the research outcomes from this project to the industry and research communities and benefit society.The goal of this project is to understand the security and privacy challenges in achieving cooperative perception among autonomous vehicles, and then design an effective and efficient privacy-preserving CNN for processing encrypted sensor data, shared from multiple vehicles. Leveraging the additive secret sharing technique, sensor data is first randomly split and encrypted into two ciphertexts, and then processed in the encrypted format by two non-colluding edge servers. To process the encrypted data, edge servers make use of a privacy-preserving CNN model that securely implements all layers in the original CNN, including the secure convolutional layer, secure activation layer, secure pooling layer, secure region proposal network layer, and secure full-connection layer. As an edge server possesses only a partial view of the sensor data, sensitive information contained in the data is protected. To optimize the privacy-preserving CNN model, cross-layer optimization and field programmable gate arrays (FPGA)-based acceleration techniques are designed to increase the speed of object detection and classification while keeping the energy consumption low. The proposed system will serve as a convincing proof-of-concept for data privacy protection using CNNs, thus opening the door to widespread adoption of privacy-preserving CNN in processing sensor data. The source code and the privacy-preserving CNN model, along with its training and testing datasets, will be made publicly available, serving as a catalyst for enabling innovative research on data privacy protection in vehicular edge systems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.