SBIR Phase I: Open Machine Learning Competitions with Private Data

这个小企业创新研究（SBIR）第一阶段项目的更广泛影响将是扩大对人工智能（AI）人才的获取，并刺激创新，以解决难题，同时保护隐私。机器学习和人工智能正在为政府、私营公司和社会部门组织带来转型变革。然而，在未来几年，创新将受到人工智能人才有限的限制。开放式创新，如机器学习（ML）竞赛，为政府和企业提供了利用全球人才库解决一些最紧迫和最棘手挑战的能力。然而，目前运行这些比赛存在巨大的障碍：数据必须提供给参与者，如果相关数据过于敏感，由于对隐私，安全或机密性的担忧而无法发布，则可能会妨碍比赛的运行。随着对数据人才的需求越来越高，政府机构、公司和其他机构都表现出了以这种方式投资的意愿。拟议的项目开发了一种大规模维护数据隐私的方法。小企业创新研究（SBIR）第一阶段项目将开发一个端到端的竞争系统，为用于构建众包算法解决方案的数据提供隐私保证。开放式ML挑战通常通过为参与者提供训练数据来学习底层模式，然后在未标记的测试数据上评估结果预测。对于许多重要的问题，以这种方式提供训练数据违反了对隐私的担忧或可能被滥用。关键的差距是保护训练数据的隐私，同时使参与者能够构建可以从中学习的模型。该项目将汇集隐私保护数据分析中最有前途的三种方法的最新进展：同态加密，联邦学习和差分隐私。每种技术都将在一个专门的挑战结构中开发和测试，该结构具有两个核心属性：1）保护敏感数据的隐私; 2）确保参赛者能够在比赛期间获得对提交模型的反馈，以告知算法改进。每个竞赛系统将产生一套性能指标，包括基准算法性能和数据隐私保证，以评估系统的可行性。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will be to expand access to artificial intelligence (AI) talent and spur innovation to solve hard problems while protecting privacy. Machine learning and AI are bringing transformational change to governments, private companies, and social sector organizations. Yet in the coming years, innovation will be hamstrung by limited access to AI talent. Open innovation, such as machine learning (ML) competitions, provides governments and firms the ability to tap into a global talent pool to solve some of their most pressing and vexing challenges. Yet there is currently an immense barrier to running these competitions: the data must be made available to participants, which can preclude running a competition if the associated data are too sensitive to release due to concerns about privacy, security, or confidentiality. With data talent in increasingly high demand, government agencies, companies, and others have demonstrated a willingness to invest in this fashion. The proposed project develops a method to maintain data privacy at scale. This Small Business Innovation Research (SBIR) Phase I project will develop an end-to-end competition system that provides privacy guarantees for data used to build crowdsourced algorithmic solutions. Open ML challenges typically work by providing participants with training data to learn underlying patterns, then evaluating resulting predictions on unlabeled test data. For many important problems, making training data available in this way violates concerns about privacy or enables abuse. The critical gap is preserving the privacy of training data while enabling participants to build models that can learn from it. This project will bring together recent advances in three of the most promising approaches in privacy-preserving data analysis: homomorphic encryption, federated learning, and differential privacy. Each technique will be developed and tested in a dedicated challenge structure with two core properties: 1) to preserve the privacy of sensitive data; and 2) to ensure competitors are able to get feedback on submitted models during the competition to inform algorithm improvements. Each competition system will result in a set of performance measures, including benchmarked algorithm performance and data privacy guarantees, to assess system feasibility.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.