Relation Mining Under Local Differential Privacy
Kai Dong, Xinwen Fu
33rd USENIX Security Symposium · Day 1 · USENIX Security '24
In an era where centralized institutions amass vast quantities of data, data mining has become an indispensable tool for extracting immense value across diverse sectors, from market analysis and social media optimization to healthcare risk prediction and fraud detection. This process transforms raw datasets into actionable insights, fueling innovation and informing critical decisions. However, the pervasive collection and analysis of sensitive information introduce significant privacy risks, particularly when central servers may not be entirely trustworthy. To address these concerns, **Local Differential Privacy (LDP)** has emerged as a robust standard for data protection, offering strong mathematical guarantees that ensure individual privacy even against adversaries with extensive background knowledge.
AI review
This talk presents LDPRM, a groundbreaking framework that finally enables robust relation mining under Local Differential Privacy. It brilliantly overcomes the "curse of dimensionality" using novel dimensionality reduction techniques like SVD and low-rank approximation, filling a critical gap in privacy-preserving data analysis and offering significant practical utility for extracting complex insights from sensitive data.