CAREER: Bio-Inspired Sensory Interfaces Incorporating Embedded Classification and Encryption

无处不在的传感和计算导致大数据分析的快速增长，可能会改变世界。这一愿景为无处不在的感知界面带来了新的挑战，以实现始终在线功能、快速信息分析和安全设计，以防止网络攻击。然而，与此同时，运行机器学习和复杂的密码算法将消耗大量电力。在将分类器和安全措施整合到传感器以实现持续监测方面也存在重大挑战。该项目提出了一个研究、教育和推广的综合计划，以开发具有理论、算法和体系结构的低功耗传感器系统，以实现传感器内的智能和安全。这项研究项目的变革性方面包括对仿生计算的基本理解，发现有用的内在设备特征，使用自适应机器学习分析实时数据，以及探索高效加密的混沌行为。这项研究将通过泛在传感和计算的发展，对社会对安全和持续实时监测的需求产生重大影响，以改善健康、交通和环境。该项目还纳入了一项综合教育计划，以鼓励和激励具有不同背景的年轻一代，特别是妇女和代表性不足的少数群体，在科学、技术、工程和数学(STEM)领域接受教育。该计划将向本科生和研究生介绍安全无处不在传感和计算的概念，并通过令人信服的例子说明基础电子和数学的易于理解的概念，为当地K-12学生创建强有力的外展活动。该项目的目标是开发超低功耗传感接口，将自主传感、分类和安全措施集成到单一硬件平台中。受生物启发的分类器结合了组合的内在特征，模拟了复杂的生物系统，其中感知、学习和决策是通过非线性和自适应模拟计算进行的。该体系结构以快速再生为驱动，提取相对时序信息，用于分层分类。该方法不再采用线性放大和精细集成，而是在时域中利用器件固有的失配和非线性来实现低电源电压下的节能计算。为了处理传感器中的实时数据，采用伯努利变分分布来逼近后验分布，利用贝叶斯方法建立了计算高效的多层神经网络。该算法将医学知识和统计分析结合到训练过程中，以适应输入信号。该算法在样本空间和特征空间中都探索了最大稀疏性，其中模型中包含了硬件约束的正则化以确保稳健性。此外，为了在传感器中执行加密，信息将被随机化为确定性噪声进行传输。利用时变映射对流水线混沌系统进行训练，可以在不产生可观测模式的情况下增强安全性，从而对抗旁路攻击。该变换函数具有物理上不可克隆的组合内在特征，以确保完整的安全措施。这确保了从边缘传感器到云的多层安全方案的数据完整性和基本身份验证，同时分类算法在传感器中本地执行，以实现无线通信的快速分析和数据缩减。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Ubiquitous sensing and computing, leading to rapid growth of big data analysis, will potentially transform the world. That vision creates new challenges for pervasive sensory interfaces to enable the always-on feature, rapid analysis of information, and design for security to prevent cyberattacks. In the meantime, however, significant power will be consumed to run machine learning and complex cryptography algorithms. Critical challenges also exist in integrating classifiers and security measures into sensors to enable continuous monitoring. This project proposes an integrated program of research, education, and outreach to develop low-power sensory systems with theory, algorithms and architectures to enable in-sensor intelligence and security. The transformative aspects of this research project include fundamental understanding of bio-inspired computing, discovering useful intrinsic device characteristics, analysis of real-time data with adaptive machine learning, and exploring chaotic behaviors for efficient encryption. This research will have a significant impact on the needs of society for secure and continuous real-time monitoring to improve health, transportation, and environment through the developments of ubiquitous sensing and computing. This project also incorporates an integrated education plan to inspire and motivate younger generations with diverse backgrounds, in particular women and underrepresented minorities, to pursue education in Science, Technology, Engineering and Mathematics (STEM) fields. The plan will introduce the concepts of secure ubiquitous sensing and computing to undergraduate and graduate students, and create strong outreach activities to local K-12 students by illustrating easily-understood concepts of fundamental electronics and mathematics with compelling examples.The goal of this project is to develop ultra-low-power sensory interfaces that integrate autonomous sensing, classification, and secure measures into a single hardware platform. The bio-inspired classifiers incorporating combinatorial intrinsic characteristics emulate sophisticated biological systems where sensing, learning, and decision making are carried out through nonlinear and adaptive analog computing. The proposed architecture is driven by fast regeneration to extract relative timing information for hierarchical classification. Instead of using linear amplification and fine integration, inherent device mismatch and nonlinearity are exploited in time domain to achieve energy-efficient computation under low supply voltages. To process real-time data in sensors, Bernoulli variational distributions are employed for approximating the posterior to develop a computationally-efficient multi-layer neural network with Bayesian methods. The algorithm integrates medical knowledge and statistical analysis into the training process for adaptation to incoming signals. The proposed algorithm explores maximum sparsity in both sample and feature spaces, where regularizations of hardware constraints are included in the model to ensure robustness. Moreover, to perform encryption in sensors, the information will be randomized into deterministic noise for transmission. The pipeline chaotic system can be trained with time-varying maps to enhance the strength of the security without creating observable patterns to counter side-channel attacks. The transformation function is built with combinatorial intrinsic characteristics, which are physically unclonable to ensure complete security measures. This ensures data integrity and basic authentication for multi-layer security schemes from the edge sensors to the cloud while classification algorithms are performed locally in sensors to achieve rapid analysis and data reduction for wireless communications.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.