Pandemic Response Modelling with Privacy Enhancing Technology: a place-centric approach

这一挑战寻求开发隐私增强技术(PETS)，该技术可以解锁卫生保健提供者之间的敏感数据共享，并使用联合学习方法来模拟感染风险。传统的机器学习技术可能会产生嵌入敏感数据的模型。联合学习方法难以将发生在单个节点中的弱信号的学习推广到聚合模型。疫情预测依赖于从敏感数据中的微弱信号进行推断，这使得问题具有挑战性。我们的方法将侧重于基于个人活动和位置模式的预测。我们建议将隐私负担转移到功能工程阶段。这使得联合学习可以使用更简单的方法，只需要处理隐私保护功能，而不是自己提供隐私保护。我们将通过应用两个关键技术来实现这一点：首先通过同态加密，这是一种允许对加密数据执行计算的密码技术。因此，在节点和中央服务器之间传递的数据是加密的。其次，我们将在系统中引入噪音，以掩盖个人的影响。添加噪声的目的是阻止发现个人数据的尝试，这个概念在数学上被定义为差异隐私。我们将采取以地点为中心的方法来应对这一挑战。我们将重点关注感染者的位置。虽然我们的方法是在地理空间数据方面实现的，但它可以简单地适用于基于其他图形数据的特征，例如，人口联系图而不是地点访问图。

This challenge seeks to develop Privacy Enhancing Technologies (PETs) that may unlock the sharing of sensitive data between health care providers, and use a federated learning approach to model infection risk. Traditional machine learning techniques can result in a model that embeds sensitive data. Federated learning approaches struggle to promote learning from weak signals that occur in a single node to the aggregate model. Pandemic forecasting relies on inferring from weak signals within sensitive data, making the problem challenging.Our approach will focus on prediction based on individuals' activity and location patterns. We propose shifting the privacy burden to the feature engineering stage. This allows the use of simpler approaches to federated learning that need only handle privacy preserving features, rather than provide privacy preservation themselves.We will do this by applying two key techniques: first through Homomorphic encryption, a cryptographic technique that allows computations to be performed on encrypted data. So that data passed between nodes and a central server are encrypted. Second, we will introduce noise into the system to mask the impact of an individual. The purpose of adding noise is to thwart attempts to discover data about an individual, a concept mathematically defined as _differential privacy_. We will take a place-centric approach to this challenge. The features that we will focus on will be the locations of infected individuals. While our approach is implemented in terms of geospatial data, it could trivially be adapted to produce features based on other graph data, for example, the population contact graph rather than the place visitation graph.