Deriving Private Information from Arbitrarily Projected Data

Distance-preserving projection based perturbation has gained much attention in privacy-preserving data mining in recent years since it mitigates the privacy/accuracy tradeoff by achieving perfect data mining accuracy. One apriori knowledge PCA based attack was recently investigated to show the vulnerabilities of this distance-preserving projected based perturbation approach when a sample dataset is available to attackers. As a result, non-distance-preserving projection was suggested to be applied since it is resilient to the PCA attack with the sacrifice of data mining accuracy to some extent. In this paper we investigate how to recover the original data from arbitrarily projected data and propose AKICA, an Independent Component Analysis based reconstruction method. Theoretical analysis and experimental results show that both distance-preserving and non-distance-preserving projection approaches are vulnerable to this attack. Our results offer insight into the vulnerabilities of projec...

• Type: Conference
• Year: 2007
• Where: PAKDD
• Authors: Songtao Guo, Xintao Wu
• Keywords: Data Mining, Data Mining Accuracy, PAKDD 2007, Privacy-preserving Data Mining