CDS&E: Collaborative Research: Private Data Analytics, Synthesis, and Sharing for Large-Scale Multi-Modal Smart City Mobility Research

CDS

• 批准号: 2002985
• 负责人: David Evans
• 金额: $ 16.5万
• 依托单位: University of Virginia Main Campus
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-07-01 至 2023-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2002985&HistoricalAwards=false
• 关键词: CDS& Collaborative; Research; Private; Data

Given the trend towards urbanization, understanding real-time human mobility in urban areas has become increasingly important for many research areas from Mobile Networking, to Transportation/Urban Planning, Behavior Modeling, Emergency Response, to recent Pandemic Mitigation. Many analytical models have been proposed to understand human mobility based on mobility data. However, most of these data are proprietary and cannot be accessed by the research community at large. Fortunately, based on the latest expansion of urban infrastructures, such mobility data has been collected by city government agencies and some companies that are willing to share the data for social good. However, a key challenge is the privacy concern since such data usually have sensitive information and system design details for potential privacy and security issues. To address this issue, the project aims to generate realistic yet synthetic mobility data through machine learning based on the real mobility data analytics and then share these realistic synthetic data with the research community. The objective of the project is to lower the entry barriers for interdisciplinary researchers in mobility data-intensive research aimed at addressing major scientific/societal challenges related to urban mobility.The core merit of the project lies in integrating two aims, i.e., privacy-preserving data synthesis and data integration, for large-scale smart city mobility research. For the first research aim, the project plans to utilize recent advances in Generative Adversarial Networks (GANs) to enable large-scale mobility data synthesis. The goal is to achieve the individual-level release of realistic synthetic mobility data by GAN-based models targeting key characteristics of human mobility. The GAN architecture proposed has novel technical components to augment basic GAN frameworks, which optimize the fundamental trade-off between privacy (regarding removing/obfuscating sensitive mobility features) and utility (in terms of preserving non-sensitive mobility features) with long-range dependencies (in terms of repeated mobility patterns) revealed. For the second research aim, the PIs plans to perform multi-modal data integration based on aligned multi-tensor decomposition under mobility semantics. The technical approach proposed is to enable multi-modal data integration based on synthetic single-modal data for comprehensive mobility modeling with a set of machine learning techniques including novel mobility semantic learning and multi-tensor decomposition with aligned spatiotemporal granularity.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.

鉴于城市化的趋势，了解城市地区的实时人员流动对于从移动网络到交通/城市规划、行为建模、应急响应到最近的流行病缓解等许多研究领域变得越来越重要。人们提出了许多分析模型来基于流动性数据来理解人类流动性。然而，这些数据大部分都是专有的，整个研究界无法访问。幸运的是，基于城市基础设施的最新扩建，此类出行数据已被市政府机构和一些愿意为了社会公益而共享数据的公司收集。然而，一个关键的挑战是隐私问题，因为此类数据通常包含敏感信息和潜在隐私和安全问题的系统设计细节。为了解决这个问题，该项目旨在通过基于真实移动数据分析的机器学习生成真实且合成的移动数据，然后与研究界共享这些真实的合成数据。该项目的目标是降低跨学科研究人员进行移动数据密集型研究的进入门槛，旨在解决与城市移动相关的重大科学/社会挑战。该项目的核心优点在于整合两个目标，即隐私保护数据合成和数据集成，用于大规模智慧城市移动研究。对于第一个研究目标，该项目计划利用生成对抗网络（GAN）的最新进展来实现大规模移动数据合成。目标是通过基于 GAN 的模型针对人类流动性的关键特征，实现个人层面的真实合成流动性数据的发布。提出的 GAN 架构具有新颖的技术组件来增强基本的 GAN 框架，从而优化隐私（关于删除/混淆敏感移动性特征）和实用性（关于保留非敏感移动性特征）与远程依赖性之间的基本权衡（就重复的移动模式而言）揭示了。对于第二个研究目标，PI 计划基于移动语义下对齐的多张量分解来执行多模态数据集成。所提出的技术方法是基于合成单模态数据实现多模态数据集成，通过一套机器学习技术进行全面的移动建模，包括新颖的移动语义学习和具有对齐时空粒度的多张量分解。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

CryptGPU: Fast Privacy-Preserving Machine Learning on the GPU

CryptGPU：GPU 上的快速隐私保护机器学习

• DOI: 10.1109/sp40001.2021.00098
• 发表时间: 2021-05
• 期刊: IEEE Symposium on Security and Privacy
• 作者: Tan, Sijun;Knott, Brian;Tian, Yuan;Wu, David J.
• 通讯作者: Wu, David J.

Towards Return Parity in Markov Decision Processes

马尔可夫决策过程中的回报平价

• 发表时间: 2022-01
• 期刊: 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022
• 作者: Chi, J.;Shen, J.;Dai, X.;Zhang, W;Tian, Y.;Zhao, H.
• 通讯作者: Zhao, H.

Understanding and Mitigating Accuracy Disparity in Regression

了解并减少回归中的准确性差异

• 发表时间: 2021-01
• 期刊: Thirty-eighth International Conference on Machine Learning (ICML 2021
• 作者: Chi, J;Tian, Y;Gordon, G;Zhao, H.
• 通讯作者: Zhao, H.

TransRisk: Mobility Privacy Risk Prediction based on Transferred Knowledge

TransRisk：基于转移知识的移动隐私风险预测

• DOI: 10.1145/3534581
• 发表时间: 2022-07
• 期刊: Wearable and Ubiquitous Technologies
• 作者: Xie, Xiaoyang;Hong, Zhiqing;Qin, Zhou;Fang, Zhihan;Tian, Yuan;Zhang, Desheng
• 通讯作者: Zhang, Desheng

Model-Targeted Poisoning Attacks with Provable Convergence

具有可证明收敛性的模型目标中毒攻击

• 发表时间: 2021-01
• 期刊: Proceedings of the 38 th International Conference on Machine Learning
• 作者: Suya, F.;Mahloujifar, S;Suri, A;Evans, D;Y. Tian
• 通讯作者: Y. Tian