CNS Core: Small: Privacy-Preserving On-Device Intelligence in the IoT Era

设备上的人工智能(AI)在物联网(IoT)时代发挥着越来越重要的作用。作为一个关键的推动因素，模型压缩可以在有限的物联网资源中容纳最先进的AI模型，而不会显著影响预测性能。鉴于物联网设备之间的连接，设备上的智能预计不仅可以更快、更安全地为个别物联网用户提供服务，还可以支持更广泛的物联网社区。然而，不断增长的物联网市场使物联网设备成为对手的主要目标。与此同时，最近的研究表明，人工智能模型对隐私构成了严重威胁。当物联网遇到人工智能时，设备上的智能更有可能受到多个新的攻击载体的影响，具有重大的潜在影响。该项目包括三个研究推力，以发展物联网时代的隐私保护设备上智能：(I)探索新的漏洞和相应的对策，以私密的设备上推理模型压缩；(Ii)建立一个私人协作推理框架，以保护高预测性能的中间设备上表示；和(Iii)调查不确定性感知的知识聚合，轻量级，设备上混淆，以保护协作训练隐私。拟议的研究将通过模拟工具、大规模试验台和真实世界部署进行严格评估。该项目的成功将给物联网时代的一系列领域带来实实在在的好处，包括医疗保健、网络物理系统、交通和教育。拟议的研究将释放智能物联网应用和服务的潜力，同时保护隐私，为更多样化的社区实现人工智能民主化。项目中开发的程序、模型和试验台将公开发布，以推动社区在这一研究领域的进步。此外，研究成果将被整合到现有的课程、教育计划和K-12推广活动中，以鼓励学生，特别是代表不足的群体，继续进行STEM教育和研究。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。

On-device artificial intelligence (AI) is playing an increasingly important role in the Internet-of-Things (IoT) era. As a key enabler, model compression can accommodate state-of-the-art AI models within limited IoT resources without significantly compromising prediction performance. Given the connectivity among IoT devices, on-device intelligence is expected to not only serve individual IoT users more quickly and safely, but also support the broader IoT community. However, the growing IoT market makes IoT devices a primary target of adversaries. Meanwhile, recent research has shown severe privacy threats of AI models. When IoT meets AI, on-device intelligence is more likely to be subject to multiple new attack vectors, with significant potential impacts. This project consists of three research thrusts to develop privacy-preserving on-device intelligence in the IoT era: (i) exploring new vulnerabilities and the corresponding countermeasures for private on-device inference with model compression; (ii) establishing a private collaborative inference framework to protect intermediate on-device representations with high prediction performance; and (iii) investigating uncertainty-aware knowledge aggregation with lightweight, on-device obfuscation to protect collaborative training privacy. The proposed research will be rigorously evaluated via simulation tools, a large-scale testbed, and real-world deployment.The success of this project will bring tangible benefits to a range of domains in the IoT era, including healthcare, cyber-physical systems, transportation, and education. The proposed research will unlock the potential of intelligent IoT applications and services while preserving privacy, democratizing AI for a more diverse community. The programs, models, and testbeds developed in the project will be publicly released to boost community progress in this research area. Further, research findings will be integrated within existing courses, educational programs, and K-12 outreach activities to encourage students, especially underrepresented groups, to pursue STEM education and 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.