Ultra Low Power Secure Processors for Emerging Applications at the Edge

处理器支撑着我们的计算能力。用户需要更大数量、更高效和更安全地访问、操作和分析数据。增强型处理器设计是架构创新和技术改进的结合。摩尔定律（每两年单位面积的晶体管数量增加一倍）和Dennard缩放（晶体管缩小时保持恒定的功率密度）推动了基础晶体管技术，在过去50年中产生了更好的处理器。不幸的是，由于物理限制，Dennard缩放已不再适用，摩尔定律也将很快结束。鉴于底层设备技术的限制越来越多，架构师的负担是提供更高的性能和更高的能源效率。与此同时，安全性和隐私性已成为处理器设计的首要问题; SPECTRE等头条漏洞突出了对更安全架构的迫切需求。 拟议的研究通过安全的近似计算创新来解决这些迫切的需求，为运行机器学习等需要安全和隐私保护机制的应用程序的边缘和物联网（IoT）设备设计更好的处理器。这些都是越来越重要的领域：机器学习正在引领下一次计算机革命，到2025年，物联网设备预计将超过800亿，用于许多领域（例如，卫生、金融、先进制造业、交通和通信）。 近似计算范例源于这样一种观察，即并非所有应用都需要精确计算来产生可接受的结果。应用程序服从近似共享至少三个特征之一：噪声输入，统计计算，或容忍不精确。机器学习应用程序和在噪声传感器数据上运行的物联网设备都属于这种范式。架构师可以利用更软的正确性要求来换取更高的性能和/或更低的能耗。拟议的研究利用近似计算来开发两个目标：1）创新的近似计算技术，专门增强安全性和隐私性; 2）新的架构，以超低功耗和面积开销为目标，针对能量收集物联网设备提供安全性和隐私。 PI将通过在HQP培训中强调EDI原则，通过PI支持EDI的专业服务以及通过研究本身的基本基础来解决公平，多样性和包容性（EDI）。分析包含代表性不足的个人信息的数据集可能会泄露敏感信息。 需要在数据中加入噪音，以掩盖个人的贡献。通过设计安全的体系结构，以帮助在不同的隐私，EDI的关注是在这个研究议程的最前沿。

Processors underpin our ability to compute. Users need to access, manipulate and analyze data in greater quantities, more efficiently and with better security. Enhanced processor designs are a combination of architectural innovations and technology improvements. Moore's Law (which doubles the number of transistors per unit area every 2 years) and Dennard scaling (which maintains constant power density as transistors shrink), have driven the underlying transistor technology that has resulted in better processors over the last 50 years. Unfortunately, due to physical limitations, Dennard scaling has ceased to apply and Moore's Law will end soon. Given the growing limitations to the underlying device technology, the burden is on architects to deliver increased performance and greater energy efficiency. At the same time, security and privacy have become first order processor design concerns; headline-making vulnerabilities such as SPECTRE highlight the dramatic need for more secure architectures. The proposed research addresses these pressing needs through secure approximate computing innovations to design better processors for edge and internet of things (IoT) devices running applications such as machine learning that require security and privacy-preserving mechanisms. These are increasingly critical domains: machine learning is leading the next computer revolution, and IoT devices are projected to number more than 80 billion by 2025, with uses in many sectors (e.g., health, finance, advanced manufacturing, transportation and communication). The approximate computing paradigm stems from the observation that not all applications require precise computation to produce an acceptable result. Applications amenable to approximation share at least one of three characteristics: noisy input, statistical computations, or a toleration of imprecision. Both machine learning applications and IoT devices that operate on noisy sensor data fall into this paradigm. Architects can leverage softer correctness requirements to trade accuracy for increased performance and/or reduced energy consumption. The proposed research leverages approximate computing to develop two objectives: 1) innovative approximate computing techniques that specifically enhance security and privacy and 2) new architectures that provide security and privacy with ultra-low power and area overheads targeting energy-harvesting IoT devices. The PI will address Equity, diversity and inclusion (EDI) by emphasizing EDI principles in HQP training, through the PI's professional service supporting EDI, and through the fundamental underpinnings of the research itself. Analyzing data sets containing information about under-represented individuals can potentially reveal sensitive information. Noise needs to be added to the data to obscure an individual's contribution. Through the design of secure architectures to aid in differential privacy, EDI concerns are at the forefront of this research agenda.