Collaborative Research: SaTC: CORE: Small: Securing Brain-inspired Hyperdimensional Computing against Design-time and Run-time Attacks for Edge Devices

许多计算应用程序依赖于机器学习（ML）算法来分析数据中的模式，并对遇到的新数据进行预测。这些机器学习分类器的许多最新进展使用了基于神经网络的方法；然而，神经网络通常需要大量的数据、内存和处理能力。近年来，受大脑启发的超维计算（HDC）作为一种资源较少的构建分类器的方法出现了，这种方法非常适合于计算能力较弱的小型计算设备。然而，就像其他ML分类器架构一样，HDC模型可能会受到攻击者的威胁，攻击者想要降低模型的性能，插入后门“触发器”，让攻击者通过提供秘密输入来控制设备，或者窃取模型本身。但是，对HDC模型中的这些安全风险的研究，并没有像对HDC性能的研究那样深入。本项目的目标是通过更好地理解HDC安全漏洞和防御来缩小这一差距。这包括分析HDC模型可能受到攻击的空间，将神经网络中的攻击和防御与HDC模型中的攻击和防御进行比较，并开发与HDC模型本身一样有效、高效和轻量级的防御，以便它们也可以部署在计算能力有限的设备中。该项目通过系统地研究HDC计算的攻击面，从设计到运行，从算法到硬件，为基于HDC的边缘设备推理铺平了道路。首先，探讨了与HDC相关的漏洞，并系统地定义了其独特的攻击面。因此，它研究了来自对抗性输入、模型扰动和逆向工程的HDC模型性能和隐私的关键威胁。其次，它通过结合算法，硬件和系统级方法探索有效和高效的防御策略。提出的工作中的一个关键见解和工具是将基于神经网络的模型和HDC模型相关联的方法；这将允许比较研究，以及开放的可能性，以适应现有的攻击和防御基于神经网络的架构，以HDC环境。这些科学成果将有助于重塑支持hdc的计算系统，使其更安全、更健壮。该项目也包含了重要的教育内容，并提供了大量的机会来培养和吸引来自弱势群体的学生从事计算机科学和计算机科学研究。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。

Many computing applications depend on machine learning (ML) algorithms that analyze patterns in data and make predictions about new data they encounter. Many recent advances in these machine learning classifiers use approaches based on neural networks; however, neural networks often require large amounts of data, memory, and processing power. Brain-inspired hyperdimensional computing (HDC) has emerged in recent years as a less resource-heavy approach to building classifiers that are well-suited for smaller computing devices that have less computing power. However, just like other ML classifier architectures, HDC models may be threatened by attackers who want to degrade the models' performance, insert backdoor "triggers" that let attackers take control of devices by presenting secret inputs, or steal the models themselves. However, these security risks in HDC models are not as well-studied as HDC performance. This project's goal is to close that gap through a better understanding of HDC security vulnerabilities and defenses. This includes analyzing the space of possible attacks on HDC models, drawing parallels between attacks and defenses in neural networks and those in HDC models, and developing defenses that are as effective, efficient, and lightweight as the HDC models themselves so they can too be deployed in devices with limited computing power.This project paves the way for HDC-based inference on edge devices by systematically investigating the attack surface for HDC computing, from design time to run time and from algorithm to hardware. First, it explores the vulnerabilities associated with HDC and systematically defines its unique attack surface. Accordingly, it investigates critical threats against HDC model performance and privacy from adversarial input, model perturbation, and reverse engineering. Second, it explores effective and efficient defense strategies by incorporating algorithmic-, hardware-, and system-level methods. A key insight and tool in the proposed work are methods for relating neural network-based models and HDC models; this will allow for comparative studies as well as open possibilities for adapting existing attacks and defenses on neural network-based architectures to HDC contexts. The scientific outcomes will help reshape HDC-enabled computing systems toward greater security and robustness. The project also contains a significant educational component and provides abundant opportunities to nurture and attract students from under-represented groups to engage in computer science and computer science research.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.