Hardware Security for Approximate Computing

随着IBM的2nm芯片将摩尔定律推向极限，传统的计算技术正在努力在功耗限制下提供高性能计算。受人脑容错能力的启发，近似计算可以在不影响某些人类感知和识别相关计算的结果（如准确性）的情况下大幅降低计算机功耗，这些计算只要求结果是近似的，而不是准确的。例子包括人工智能（AI）、深度学习（DL）、图像处理，甚至一些加密方案。然而，近似计算已被证明具有安全漏洞，因为固有错误的不可预测性可能与恶意修改无法区分。由于近似计算固有的功率和面积节省，安全对策也应该是轻量级和高效的。因此，本提案的目的是使用先进的硬件安全技术来开发具有最佳安全保护和最佳系统效率的近似计算技术。目前，尚未对近似计算的安全性或保护此类设计的对策进行全面的研究。物理不可克隆函数（PUF）作为一种轻量级硬件安全原语，是保护资源受限应用程序（如近似计算）的最佳候选方法之一。PUF可用于根据硅芯片的制造工艺变化为电子设备生成唯一的数字指纹。目前，在传统计算中对PUF进行了广泛的研究，但目前还没有采用近似技术的有效的内在PUF设计。这个项目是及时的，因为近似计算已经迅速引起学术界和工业界的关注，因为它解决了计算系统的基本障碍之一，功耗，但它也开辟了新的攻击载体。这个项目将开发一个基于近似处理器正常操作的内在PUF设计，而不需要额外的硬件资源。该项目还将首次解决如何实现安全有效的近似计算设计。泰雷兹英国公司是为国防、安全、航空航天和运输部门设计和建造关键任务信息系统的领导者，该公司已经邀请PI加入泰雷兹CyRes-Advance项目，通过考虑硬件安全来研究联网和自动驾驶车辆（cav）的安全保护。泰雷兹将提供25万英镑的实物支持，如硬件设计的技术咨询/审查，使用泰雷兹CAV测试平台和项目的实验验证，以加快研究进程并产生高质量的研究成果。

As IBM's 2nm chip is pushing Moore's law approaching its limit, conventional computing techniques are struggling to offer high performance computing within power consumption constraints. Inspired by the fault tolerance capability of the human brain, approximate computing, which is error tolerant, can offer a huge reduction in computer power consumption without affecting the results (such as accuracy) of certain human perception and recognition related computation that only require a result to be approximate, rather than accurate. Examples include Artificial Intelligence (AI), Deep Learning (DL), image processing and even some cryptographic schemes. However, approximate computing has been shown to have security vulnerabilities due to the unpredictability of intrinsic errors that may be indistinguishable from malicious modifications. Due to the inherent power and area savings achieved by approximate computing, security countermeasures shold also be lightweight ande efficient. Hence, the aim of this proposal is to use advanced hardware security techniques to enable the development of approximate computing technologies that have both optimal security protection and optimal system efficiency. Currently, no comprehensive research has been conducted to date into security of approximate computing or into countermeasures that protect such designs.Physical unclonable function (PUF), as a lightweight hardware security primitive, is one of the best candidates for securing resource-constrained applications, such as approximate computing. A PUF can be used to generate a unique digital fingerprint for an electronic device based on manufacturing process variations of silicon chips. Currently, PUFs have been widely studied for conventional computing but no effective intrinsic PUF designs using approximate techniques have been presented. This project is timely because approximate computing has rapidly attracted attention from both academica and industry, as it addresses one of the fundamental barriers in computing systems, power dissipation, but it has also opened new vectors of attacks. This project will develop an intrinsic PUF design based on the normal operations of an approximate processor without the need for addtional hardware resource. The project will aslo address for the first time how to achieve secure and effective approximate computing designs.Thales UK, a leader in designing and building mission-critical information systems for the defence, security, aerospace, and transportation sections, has already invited the PI to join the Thales CyRes-Advance project to investigate security protection for connected and autonomous vechicles (CAVs) by considering hardware security. Thales will provide £250k in-kind support, such as technical advice/review of the hardware design, access to Thales CAV test platform and experimental validation for the project, to accelerate the research process and produce high-quality research outputs.