SaTC: CORE: Medium: Collaborative: Privacy Attacks and Defense Mechanisms in Online Social Networks

点击翻译按钮获取中文摘要

In online social networks, people and their connections often share personal information, such as demographics, interests, and opinions, and leave traces of their interaction with others and content in the network. Not everyone wants to share personal information; however, people's attributes are correlated with each other among themselves, with attributes of nearby people in the network, and between a person's accounts on different networks. These correlations create risks around inferring attributes people would rather keep private. This project will try to identify and quantify the risks by developing new ways to infer attributes by leveraging these correlations, then develop defense mechanisms in two common social networking tasks. For querying social network datasets, which is commonly used in advertising and research, the researchers will develop new differential privacy techniques for networks to ensure that query results do not inadvertently identify individual users or their attributes. For matching social network profiles, which is often used in recommender systems, the team will develop novel similarity matching methods that work on encrypted personal data. Overall, the research will provide a deeper understanding of the risks of inadvertent leakage of personal information and possible technical and policy approaches for addressing those risks. The project will also provide research opportunities for both graduate and undergraduate students at three institutions, and the research team will emphasize recruiting students from historically underrepresented groups in computing both at the college and high school level. The project is organized around three main thrusts. The first thrust is to develop inference attacks on users' attributes and identity in social networks. To do this the team will first compute the relative discriminatory power of different attributes based on their distributions in the network, then use this and network structural information to perform attribute inference through affinity propagation. The second thrust focuses on improving differential privacy protection for graph queries. For this, the team will define similarity metrics that account for the non-independence of edges in social networks to better protect attribute privacy and develop new query techniques based on subgraph partitioning and consideration of the sensitivity of the query function. They will also develop new variants of differential privacy based on k-anonymity that hide a user's attributes relative to those of similar users. The third thrust explores how to do profile matching without revealing sensitive personal information, inspired by ideas from secure multiparty computation. Here, the team will develop efficient and accurate methods to do dot-product computation on data protected by chaos-based encryption and keyword search on data protected by attribute-set-based encryption, as well as hashing-based approaches to compute image similarity without sharing the image data itself. The team will release its code, suitably protected datasets, and tutorials and educational materials through a dedicated project website, and do outreach to members of underrepresented groups through the McNair programs, Women in Computer Science, the Society of Hispanic Professional Engineers, and the National Society of Black Engineers.