Towards practical, end-to-end differentially private machine learning

无处不在的机器学习（ML）驱动的应用程序需要越来越多的数据收集，使其用户面临越来越大的隐私风险。差分隐私（DP）等研究工具的存在是为了利用机器学习的力量，而不会出现隐私泄露的不当风险。然而，它们很难大规模使用，而且它们对ML算法性能的不利影响阻碍了它们的采用。该研究计划的长期目标是通过在现有的ML基础设施中构建对差分隐私的支持，并最大限度地减少其对ML模型性能的影响，从而在ML应用程序中实现差分隐私的实际端到端部署。目标、方法和HQP实现这一愿景需要对DP做出各种贡献，从新的理论工具，到新的算法和新的资源管理组件，所有这些都集成到现有的ML集群编排系统中。在接下来的五年里，我提出了三个互补的研究目标，它们将共同实现DP与现有ML基础设施的第一个实用的交钥匙集成。这些研究目标是：1。设计有效的隐私资源管理算法，使训练和更新多个ML模型的应用程序不会过度增加收集数据的隐私泄漏；2. 利用数据生成技术以保护隐私的方式在模型之间共享信息，在不增加整体隐私泄漏的情况下提高每个模型的性能；和3。创建工具以促进应用程序生成的所有ML模型的开发和审计。在“发现基金”的五年里，我计划吸引和培养一批不同的学生来应对这些挑战，其中包括两名博士生和五名硕士生。我还将指导10个本科科研项目。我将培养一个包容、支持和激发智力的实验室文化，有利于团队的学习和研究进展。每个研究目标的进展将推动最先进的隐私保护ML，并将在顶级计算机科学研究场所发表，参与增加加拿大对HQP在该领域的吸引力。该提案的预期结果是在基于ML的应用程序中部署DP的第一个端到端解决方案，将使加拿大公司、非营利组织和政府机构更容易利用ML的力量（例如在医疗保健、教育、金融领域），满足加拿大严格的隐私标准，并降低数据泄露的频率和成本。反过来，这将使加拿大人能够享受ML的承诺，而不会对他们的隐私造成不必要的风险。HQP I将接受机器学习、系统和隐私等关键技能的培训，将是帮助在加拿大建立相关能力的理想人选。

Research Program Ubiquitous Machine Learning (ML) driven applications require ever increasing amounts of data collection, exposing their users to growing privacy risks. Research tools, such as Differential Privacy (DP), exist to leverage the power of ML without undue risks of privacy leakage. However, they are difficult to use at large scale, and their adverse impact on ML algorithms' performance hinders their adoption. The long term objective of this research program is to enable practical, end-to-end deployments of Differential Privacy in ML applications, by building support for Differential Privacy into existing ML infrastructure, and minimizing its impact on the performance of the ML models. Objectives, Methodology, and HQP Realizing this vision requires a varied set of contributions to DP, from new theoretical tools, to new algorithms, and new resource management components, all integrated into existing cluster orchestration systems for ML. For the next five years, I propose three complementary research objectives that together will enable a first practical, turn-key integration of DP into existing ML infrastructure. These research objectives are: 1. designing algorithms for efficient privacy resource management, so that applications training and updating multiple ML models do not overly increase privacy leakage from the collected data ; 2. using data generation techniques to share information between models in a privacy-preserving fashion, improving each model's performance without increasing the overall privacy leakage ; and 3. creating tools to facilitate the development and audit of all ML models produced by an application. Over the five years of the Discovery Grant, I plan to attract and train a diverse group of students to tackle these challenges, including two PhD students and five Master's students. I will also supervise ten Undergraduate research projects. I will foster an inclusive, supportive, and intellectually stimulating lab culture conducive to learning and research progress as a team. Impact Progress along each research objective will advance the state-of-the-art in privacy preserving ML, and will be published in top Computer Science research venues, participating to increasing Canada's attractiveness to HQP in the domain. The anticipated outcome of this proposal, a first end-to-end solution for DP deployments in ML based applications, will make it easier for Canadian companies, non-profits, and government agencies to leverage the power of ML (e.g. in healthcare, education, finance) meeting Canada's strong privacy standards, and lowering the frequency and cost of data breaches. This will in turn allow Canadians to enjoy the promises of ML without undue risks to their privacy. The HQP I will train in the critically demanded skills of ML, systems, and privacy, will be ideal candidates to help build the relevant capabilities in Canada.