Infrastructure for Privacy-Preserving Data Analysis

数字经济是由(大数据)驱动的。几乎所有利益攸关方无处不在的数据收集都被利用来获取经济利益。例如，金融机构可以通过分析客户的信用卡交易和手机或在线数据，为客户提供更好的服务。然而，这些数据不仅可能被用于社会利益，还可能被用于不受欢迎甚至被禁止的用途，例如，侵犯个人隐私。因此，根据法律和为社会所接受，有必要最大限度地保护个人隐私，防止不必要的推断，同时仍然允许将数据用于社会或经济利益。这一挑战加剧，因为数据可能分布在几个实体之间，每个实体都被委托为特定目的提供数据。 密码学可以在数据存储、传输和处理期间保护隐私，例如使用同态加密或安全多方计算。在本提案中，我们的目标是启用并优化这些技术在图形处理单元(GPU)中的使用效率。GPU通常用于加速大数据分析，例如机器学习或推理。然而，由于它们使用浮点算法，因此如何将加密技术适应于GPU并不是显而易见的。我们的目标是设计和比较不同的软件方法，以在GPU中提供大数据分析的隐私保护处理。为了评估修改后的软件基础设施(如系统库)的效率影响，以及它们对长期运行程序的交互调试，需要由几个最先进的GPU组成的专用硬件基础设施。 这个GPU基础设施支持的研究项目覆盖了一个高度多样化的大型研究团队，并与行业合作伙伴密切合作完成。我们计划使我们在拟议的基础设施上开发的所有软件都作为开源软件提供。此外，部署加密协议和软件库后，可以保护隐私地使用其他计算基础设施(如加拿大计算)或云计算服务(如亚马逊的AWS)来处理敏感数据，如医疗保健数据、个人身份信息或商业秘密。因此，拟议的基础设施提高了相关行业研究项目的可扩展性和部署选项，包括外包云基础设施，并为总部提供了行业需求最高的IT技能--安全/隐私、云计算和数据科学--的实践培训。

The digital economy is driven by (big) data. The ubiquitous collection of data by almost all stakeholders is exploited for economic benefit. For example, financial institutions may provide better services to their customers by the analysis of their credit card transactions and cell phone or online data. However that data may not only be used for social benefit, but also unwanted or even prohibited uses, e.g., violating one's privacy. Hence, it is necessary by law and for social acceptability to protect the privacy of individuals to the utmost possible extent and prevent unwanted inferences while still enabling the use of the data for the social or economic benefit. This challenge is exacerbated, since the data may be distributed among several entities each of whom has been entrusted with the data for a specific purpose. Cryptography can protect privacy during data storage, transmission, and processing, e.g. using homomorphic encryption or secure multi-party computation. In this proposal we aim to enable and optimize the efficiency of these techniques for the use in graphics processing units (GPUs). GPUs are often used to accelerate big data analysis, e.g. machine learning or inference. However, due to their use of floating-point arithmetic it is non-obvious how to adapt encryption technologies to GPUs. We aim to design and compare different software approaches to provide privacy-preserving processing of big data analyses in GPUs. To evaluate the efficiency impact of a modified software infrastructure, such as system libraries, and their interactive debugging of long-running programs, a dedicated hardware infrastructure consisting of several state-of-the-art GPUs is necessary. The research programs supported by this GPU infrastructure cover a large research team with a high degree of diversity and are done in close collaboration with industry partners. We plan to make all our software developed on the proposed infrastructure available as open-source. Furthermore, the cryptographic protocols and software libraries when deployed enable the privacy-preserving use of other computing infrastructures, such as Compute Canada or cloud computing services such as Amazon's AWS, for the processing of sensitive data, such as health-care data, personally identifiable information or trade secrets. Hence, the proposed infrastructure enhances the scalability and deployment options of the related industry research programs including to outsourced cloud infrastructures and provides the HQP with hands-on training in the IT skills -- security/privacy, cloud computing and data science -- with the highest demand in industry.