Turning an enemy into an ally: Privacy In Machine Learning (Pri-ML)

数据具有巨大的潜力，可以为全球饥饿和气候变化等紧迫问题提供改变世界的解决方案。然而，以解决大问题所需的规模共享数据会带来隐私问题。我的长期目标是提供方法论基础，使我们能够在不侵犯隐私的情况下正确使用和共享这些潜在的有用数据。由于其可证明性，保护隐私的机器学习的最新进展主要集中在差分隐私上。然而，保护隐私需要在机器学习系统中注入噪声，从而导致隐私和准确性之间的根本权衡。当算法的输出对输入的任何变化都“敏感”时，这种权衡就变得更糟了，因为必须增加噪声的大小来隐藏这种变化。在研究计划的A部分，我提出了新的方法，可以有效地限制灵敏度。第二个挑战来自于差异隐私的可组合性：对数据的每次访问都会降低隐私保证。不同的私有数据生成通过创建一个可无限制重用的合成数据集来解决这个问题。与现有的假设某些数据分布或特定下游任务的工作不同，限制了生成数据的有用性，我在b部分提出了一个高度灵活的基于非参数核距离的框架来比较数据和合成数据分布。第三个挑战是由于法规（如通用数据保护法规）对未来算法设计的要求而产生的。在C部分，我提出了新的框架，以满足差异隐私和其他新兴概念，可解释性，因果关系和公平性。我们对它们相互作用的理论和定量理解将有助于开发同时考虑这些概念的实用算法。该研究项目有两个重要的社会影响。首先，B部分的技术确保了隐私保护，可以促进更多的数据共享，为公共利益服务。其次，正如Netflix纪录片《编码偏差》（Coded bias）所指出的，基于机器学习算法的自动决策系统极有可能对人们的生活产生负面影响。如C部分所建议的那样，将这些算法增强为隐私保护、公平和可解释的，可能会提高每个人的生活质量。该研究计划将支持13名学员，他们将学习机器学习的最新进展，成为下一代机器学习专家。作为一名女性首席调查员，我的目标是为女性学生提供更多的培训机会，促进该领域的性别平等。除此之外，我还打算包容代表性不足的群体，例如，地理上（如非洲、南美和东南亚）或种族上（土著人民）。

Data has great potential to provide world-changing solutions to pressing problems such as global hunger and climate change. However, sharing data at the scale required to tackle large problems comes with privacy concerns. My long-term goal is to provide methodological foundations that allow us to properly use and share these potentially instrumental data without privacy violations. Recent progress in privacy-preserving machine learning has centered around differential privacy, due to its provability. However, preserving privacy requires injecting noise in the machine learning system, leading to a fundamental trade-off between privacy and accuracy. The trade-off worsens when the algorithm's output is "sensitive" to any changes in its input, as the noise magnitude has to increase to hide that change. In Part A of the research program, I propose novel methods that can effectively limit the sensitivity. The second challenge arises due to the composability of differential privacy: every access to data reduces the privacy guarantee. Differentially private data generation solves this problem by creating a synthetic dataset that is reusable without limit. Unlike existing works assuming certain data distributions or particular downstream tasks, limiting the usefulness of the generated data, I propose a highly flexible nonparametric kernel-distance-based framework to compare the data and synthetic data distributions, in Part B. The third challenge arises due to the requirements for future algorithmic design imposed by regulations such as the General Data Protection Regulation. In Part C, I propose new frameworks that satisfy both differential privacy and other emerging notions, interpretability, causality, and fairness. Our theoretical and quantitative understanding of their interplay will be instrumental in developing practical algorithms that consider these notions simultaneously. There are two important societal impacts of the research program. First, the techniques like those in Part B ensure privacy protection which can promote more data sharing for the public good. Second, as the Netflix documentary Coded bias points out, automatic decision-making systems based on machine learning algorithms have a great potential to adversely affect people's lives. Augmenting those algorithms to be privacy-preserving, fair, and interpretable as suggested in Part C could potentially improve the quality of everyone's lives. The research program will support 13 trainees, who will learn about a broad range of recent advances in machine learning to be the next generation of machine learning experts. As a female primary investigator, my goal is to give more training opportunities to female students for gender equality in the field. Beyond that, I also intend to be inclusive to under-represented groups, e.g., geographically (such as Africa, South America, and Southeast Asia), or ethnically (indigenous people).