On the use of spectral filtering for privacy preserving data mining

Abstract

Randomization has been a primary tool to hide sensitive private information during privacy preserving data mining. The previous work based on spectral filtering, show the noise may be separated from the perturbed data under some conditions and as a result privacy can be seriously compromised. In this paper, we explicitly assess the effects of perturbation on the accuracy of the estimated value and give the explicit relation on how the estimation error varies with perturbation. In particular, we derive one upper bound for the Frobenius norm of reconstruction error. This upper bound may be exploited by attackers to determine how close their estimates are from the original data using spectral filtering technique, which imposes a serious threat of privacy breaches.