Collaborative Research: CIF: Medium: Harnessing Intrinsic Dynamics for Inherently Privacy-preserving Decentralized Optimization

合作研究：CIF：中：利用内在动力学实现固有隐私保护的去中心化优化

基本信息

批准号：
2106293
负责人：
Yongqiang Wang
金额：
$ 50万
依托单位：
Clemson University
依托单位国家：
美国
项目类别：
Continuing Grant
财政年份：
2021
资助国家：
美国
起止时间：
2021-06-01 至 2025-05-31
项目状态：
未结题

来源：
https://www.nsf.gov/awardsearch/showAward?AWD_ID=2106293&HistoricalAwards=false
关键词：
Collaborative Research CIF Medium Harnessing

项目摘要

Recent advances in communication and networking technologies lead to the emergence and proliferation of distributed interconnected systems such as swarm robotics, sensor networks, smart-grid, the Internet of Things, and collaborative machine-learning systems. A task that is fundamental to the operation of these systems is decentralized optimization, where participating nodes cooperate to minimize an overall objective function that is the sum (or average) of individual nodes’ local objective functions. Moreover, since individual nodes’ local objective functions may bear sensitive information of local nodes such as medical records in collaborative learning and user energy-consumption profiles in a smart grid, in many cases, the decentralized optimization algorithm has to make sure that a participant’s sensitive information is protected from being inferable by other participating nodes or external eavesdroppers. Although plenty of results have been proposed for decentralized optimization, most of these results do not consider the problem of privacy protection. Conventional information-technology privacy mechanisms are inappropriate for decentralized optimization because they either have to compromise the accuracy of optimization (in, e.g., differential-privacy-based approaches) or incur heavy extra computation/communication overhead (in, e.g., cryptography-based privacy approaches). The lack of effective privacy solutions for decentralized optimization not only severely hinders the social adoption of new technologies, but also leads to potential vulnerabilities since stealing private information is usually the basis for sophisticated cybersecurity attacks. Leveraging the iterative properties of decentralized optimization algorithms, the project aims to establish a new privacy-preserving approach for decentralized optimization that neither compromises optimization accuracy nor incurs large computation/communication overhead. Combined with the additional merit of needing no assistance of a trusted central coordinator, the proposed approach is expected to transformatively advance privacy-preservation in networked systems and make impacts in many applications ranging from connected vehicles, swarm robotics, smart grid, sensor networks, to collaborative machine learning. Leveraging control theory, this project seeks to establish methodologies and associated theories for inherently privacy-preserving decentralized optimization by exploiting the intrinsic dynamical properties of decentralized optimization. Besides maintaining optimization accuracy, the dynamics-based privacy approach is also free of encryption, which not only guarantees limited extra computation/communication overhead, but also promises a decentralized implementation without the assistance of any trusted third party or data aggregator. The main research thrusts are to: 1) Develop a privacy framework for dynamical systems that explicitly considers the iterative evolution of information in decentralized optimization; 2) Design perturbations to dynamics that enable privacy without affecting the accuracy of decentralized optimization methods for convex problems, and quantify the effects of the perturbations on convergence speed; 3) Investigate the influence of privacy design on decentralized non-convex optimization and exploit freedom in privacy design to facilitate decentralized non-convex problems; and 4) Evaluate the results using experiments on a multi-robot platform and connected vehicles.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.

通信和网络技术的最新进展导致分布式互连系统的出现和扩散，例如群体机器人技术，传感器网络，智能网格，物联网和协作机器学习系统。对于这些系统运行至关重要的任务是分散的优化，其中参与节点协调以最大程度地减少单个节点本地目标函数的总和（或平均）的总体目标函数。此外，由于各个节点的本地目标功能可能会在智能网格中提供诸如协作学习中的医疗记录和用户能量消费型材的敏感信息，因此在许多情况下，分散的优化算法必须确保参与者的敏感信息不受其他参与的节点或外部的evevere eveloppers的推断。尽管已经提出了大量的结果进行分散的优化，但大多数结果并未考虑保护隐私问题。传统的信息技术隐私机制不适合分散优化，因为它们必须损害优化的准确性（例如，基于差异性 - 基于易位性的方法），或者会损害重大的额外计算/通信架空开销（例如，加密基于基于加密基于密码的隐私方法）。缺乏用于分散优化的有效隐私解决方案不仅严重阻碍了新技术的社会采用，而且还会导致潜在的脆弱性，因为窃取私人信息通常是复杂的网络安全攻击的基础。该项目利用分散优化算法的迭代属性，旨在为分散优化建立一种新的隐私保护方法，既不会损害优化准确性，也不会造成大型计算/通信开销。再加上不需要受信任的中央协调员的帮助的额外优点，预计拟议的方法有望在网络系统中改进隐私保护，并在许多应用程序中产生影响，从连接的车辆，Swarm Robotics，Smart Grid，Sensor网络，传感器网络，到协作机器学习。利用控制理论，该项目旨在通过利用分散优化的内在动态属性来建立固有地保护隐私化优化的方法和相关理论。除了保持优化精度外，基于动态的隐私方法还没有加密，这不仅可以保证有限的额外计算/通信开销，而且还承诺在没有任何受信任的第三方或数据聚合器的帮助的情况下将进行分散的实施。主要的研究作用是：1）为动态系统开发隐私框架，该框架明确考虑了在分散优化中信息的迭代演变； 2）设计对动态的扰动，该动力学能够在不影响凸问题的分散优化方法的准确性的情况下，并量化扰动对收敛速度的影响； 3）研究隐私设计对分散的非凸优化和利用隐私设计中的自由的影响，以促进分散的非凸问题； 4）使用在多机器人平台和互联车辆上的实验评估结果。该奖项反映了NSF的法定任务，并通过使用基金会的知识分子优点和更广泛的影响审查标准来评估诚实地通过评估。