Privacy-Preserving Natural Language Processing

保护隐私的自然语言处理

• 批准号: RGPIN-2022-05197
• 负责人: Penn, Gerald
• 金额: $ 2.99万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=750055
• 关键词: Privacy; Preserving; Natural; Language; Processing

Most of our personally identifiable information is required by law to be protected by the businesses and sovereign entities that collect it. A great deal of that information is available in text and/or speech. As natural language processing becomes more precise, and increasingly more dependent upon training data from a variety of sources, the potential for compromising private information grows. Mobile phones, for example, should not offer personal information as suggested completions to character prediction tasks, even in where the completions make sense. This research programme seeks to combine research in homomorphic encryption, numerical analysis and privacy-preserving machine learning with recent advances in natural language processing to formulate protocols and algorithms for processing text and speech with guarantees that prevent the disclosure of private information. The potential applications of these results overlap with the already substantial body of literature on several tasks, e.g., biomedical NLP/speech, or security applications of speech processing, but offers a greater level of technical sophistication than would be offered by the vast majority of generic NLP approaches. This research also offers a valuable opportunity to speculate on a range of realistic, interconnected use cases that motivate techniques with broader applicability to privacy-preserving computing in general ---- techniques that have, to date, nevertheless evaded the attention of the cryptography and security communities.

法律要求我们的大多数个人身份信息受到收集该信息的企业和主权实体的保护。大量信息可以通过文本和/或语音形式获得。随着自然语言处理变得更加精确，并且越来越依赖于各种来源的训练数据，泄露私人信息的可能性也随之增加。例如，手机不应该提供个人信息作为角色预测任务的建议完成，即使这些完成是有意义的。该研究项目旨在将同态加密、数值分析和隐私保护机器学习的研究与自然语言处理的最新进展相结合，制定用于处理文本和语音的协议和算法，并保证防止私人信息泄露。这些结果的潜在应用与一些任务上已有的大量文献重叠，例如生物医学 NLP/语音或语音处理的安全应用，但提供了比绝大多数通用 NLP 方法更高水平的技术复杂性。这项研究还提供了一个宝贵的机会来推测一系列现实的、相互关联的用例，这些用例激发了对一般隐私保护计算具有更广泛适用性的技术——迄今为止，这些技术尚未引起密码学和安全社区的关注。