Private Distributed Analytics Platform for Health Data

人类健康取决于许多复杂的因素，包括环境、性别、饮食、种族和遗传变异。因此，需要来自不同来源的大量健康数据来了解不同变量对健康和疾病的影响。然而，由于政策的差异和敏感健康数据的潜在滥用，很难在不同的地理实体之间建立集中的数据存储库。因此，原始数据永远不会离开一个机构或政府的边界。例如，马尼托巴省卫生政策中心(MCHP)不向机构以外传播原始数据。因此，设计能够在保护隐私的同时执行健康数据分析的分布式框架势在必行。联合学习是一种流行的协作学习框架，可以使数据接收者能够对存储在医疗保健组织中的私人数据进行分析。联合学习允许多个数据提供商(例如医院)在不共享原始数据的情况下构建共享的机器学习模型。每个数据提供程序在其数据集上运行模型，并且仅共享中间结果。然而，最近的研究表明，在联合学习方法中传输的聚合数据(即中间结果)仍然会导致隐私攻击。因此，纳入隐私模型以阻止此类攻击，特别是针对健康数据的攻击，对于开发针对健康数据的私人联合学习框架至关重要。这项研究计划的长期目标是开发一种数据管理系统，该系统可以为不同的应用程序集成和共享敏感数据，并且可用和可扩展，并确保数据的隐私和准确性。短期目标是使用联合学习模型为健康数据开发一个私有数据分析框架。必须满足两个重要的隐私要求：(1)必须保护其数据被收集的个人的隐私，以及(2)需要确保数据提供者相对于其他提供者的隐私。我将研究新的方法，将最先进的加密和差异隐私模型结合起来，为健康数据分析开发一个私人联合学习框架。联合学习可以克服传统集中式学习方法的局限性。然而，由于数据的敏感性质以及对高数据利用率的同时需求，医疗保健部门采用联合学习的速度一直慢于其他行业。这项研究计划的成果将通过开发一个可扩展、准确和私人的联合学习框架来帮助弥合这一差距。开发的方法也将使相关研究领域受益(例如，物联网、金融和智能交通的联合学习)。除了出版物和软件分发，该研究计划还将在隐私、数据科学和机器学习技术方面培训HQP。

Human health depends on many complex factors, including environment, gender, diet, ethnicity and genetic variations. Therefore, large amounts of health data are needed from different sources to understand the impact of different variables on health and disease. However, due to the differences in policies and potential misuse of sensitive health data, it is hard to establish centralized data repositories among different geographical entities. As a result, raw data never leaves the boundary of an institution or a government. For example, Manitoba Centre for Health Policy (MCHP) does not disseminate raw data beyond the institution. Therefore, it is imperative to design distributed frameworks capable of performing health data analysis while protecting privacy. Federated learning, a popular framework for collaborative learning, can enable data recipients to conduct private data analysis stored across healthcare organizations. Federated learning allows multiple data providers (e.g., hospitals) to build a shared machine learning model without sharing the raw data. Each data provider runs the model on its dataset and shares only intermediate results. However, recent research shows that aggregated data (i.e., intermediate results) transmitted in the federated learning approach can still cause privacy attacks. Therefore, incorporating a privacy model to thwart such attacks, particularly for health data, is crucial in developing a private federated learning framework for health data. The long-term objective of this research program is to develop a data management system that can integrate and share sensitive data for different applications, that is usable and scalable, and that ensures the privacy and accuracy of data. The short-term objective is to develop a private data analytics framework for health data using the federated learning model. Two important privacy requirements have to be satisfied: (1) the privacy of individuals whose data are being collected must be preserved, and (2) the privacy of a data provider against other providers needs to be ensured. I will study novel methods to combine state-of-the-art cryptographic and differential privacy models to develop a private federated learning framework for health data analysis. Federated learning can overcome the limitations of the traditional centralized approach. However, the adoption of federated learning in the healthcare sector has been slower than other industries due to the sensitive nature of the data and the simultaneous need for high data utility. The outcomes of this research program will help to bridge this gap by developing a scalable, accurate, and private federated learning framework. The methods developed will also benefit related research areas (e.g., federated learning for IoT, finance, and smart transportation). In addition to publications and software distributions, this research program will also train HQPs in privacy, data science and machine learning technologies.