SaTC: CORE: Small: Securing Network Embedding against Privacy Attacks

许多复杂系统采取网络的形式，如社会网络、生物网络和信息网络。众所周知，在大型网络上执行数据分析是具有挑战性的。为了应对这一挑战，已经设计了许多网络表示学习(NRL)方法来学习保存网络信息的低维向量表示(嵌入)。由于它们的优势，网络嵌入，而不是原始的网络数据，要么向公众发布，要么与第三方机器学习服务提供商或其他方共享，用于下游分析。由于网络嵌入本质上捕获了原始网络的结构和性质，因此网络嵌入是否也对输入网络数据中的敏感信息进行了编码引起了人们的关注。攻击者可以对嵌入的网络发起各种隐私攻击，以推断输入的网络数据中的敏感信息。尽管在机器学习隐私方面取得了很大进展，但现有的隐私攻击主要集中在对非图形数据(表格数据、文本和图像)进行训练的模型上。NRL模型及其输出网络嵌入的隐私性在很大程度上被忽略。对网络嵌入针对隐私攻击漏洞的性质和程度的认识，以及对有效防御机制的研究都极为有限。该项目将通过一个统一的研究计划来解决网络嵌入的核心隐私问题，该计划包括探索网络嵌入的隐私漏洞，调查这些漏洞的原因，并开发严格而实用的技术来缓解这些漏洞。该项目计划解决三个基本研究问题：(I)设计三种类型的隐私攻击，即成员身份推理攻击、属性推理攻击和属性推理攻击，以从网络嵌入中推断原始网络数据中的敏感成员/属性/属性；(Ii)分析哪些类型的数据特征和模型属性影响网络嵌入的隐私漏洞，以抵御这三种攻击；以及(Iii)设计有效的防御机制来保护网络嵌入。该项目的研究成果将通过开发新课程、让不同水平的学生参与机器学习的前沿研究以及培训女性和代表性不足的学生来广泛传播。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Many complex systems take the form of networks, such as social networks, biological networks, and information networks. It is well recognized that performing data analysis on large networks is challenging. To tackle this challenge, numerous network representation learning (NRL) approaches have been designed to learn low-dimensional vector representations (embedding) that preserve the network information. Due to their advantages, network embedding, instead of the original network data, is either released to the public or shared with third-party machine learning service providers or other parties for downstream analytics. Since the network embedding inherently captures the structure and properties of the original network, it raises the serious concern whether the embedding also encodes the sensitive information in the input network data. An adversary can launch various privacy attacks on the network embedding to infer the sensitive information in the input network data. Despite significant progress in machine learning privacy, most of the existing privacy attacks mainly focus on the models trained on non-graph data (tabular data, text, and images). Privacy of NRL models and their output network embeddings are largely ignored. The understanding of the nature and extent of vulnerabilities of network embedding against the privacy attacks as well as the investigation of effective defense mechanisms are extremely limited. This project will address the core privacy issues of network embedding through a unified research program that consists of exploring the privacy vulnerabilities of network embedding, investigating the causes of these vulnerabilities, and developing rigorous, yet practical, techniques to mitigate these vulnerabilities. The project plans to tackle three fundamental research problems: (i) design three types of privacy attacks, namely membership inference attacks, attribute inference attacks, and property inference attacks, to infer the sensitive membership/attributes/properties in the original network data from network embedding; (ii) analyze which types of data characteristics and model properties impact the privacy vulnerabilities of network embedding against the three types of attacks; and (iii) design effective defense mechanisms to secure network embedding. The research outcomes of this project will be disseminated broadly through developing new courses, involving students at various levels into cutting-edge research in machine learning, and training of female and underrepresented students.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.