An Attack on the Privacy of Sanitized Data that Fuses the Outputs of Multiple Data Miners

Abstract

Abstract-Data sanitization has been used to restrict reidentification of individuals and disclosure of sensitive information from published data. We propose an attack on the privacy of the published sanitized data that simply fuses outputs of multiple data miners that are applied to the sanitized data. That attack is practical and does not require any background or additional information. We use a number of experiments to show scenarios where an adversary can combine outputs of multiple miners using a simple fusion strategy to increase their success chance of breaching privacy of individuals whose data is stored in the database. The fusion attack provides a powerful method of breaching privacy in the form of partial disclosure, for both anonymized and perturbed data. It also provides an effective way of approximating predictions of the best miner (a miner that provides the best results among all considered miners) when this miner cannot be determined.

An Attack-Resistant Hybrid Data-Privatization Method with Low Information Loss

“…We implemented all data reconstruction methods on the 4500 datasets mentioned above. In each case, we obtained an estimation of the original data set which was then compared with the original data set in terms of the success measures described in the papers [8], [9] and [10]. In each case, although we tested (the same) 4500 matrices in each attack, we present the details of only one of our matrices -one which has 6400 entries.…”

“…We use the same assumptions on data mentioned in [8] (SPF), [9] (BE-DR) and [10] (MDMF) as appropriate, and use the notation of Table 1. We generate 4500 matrices using the algorithm described in Section 2; 1500 of these were of size 400, 1500 of size 1600 and 1500 of size 6400.…”

“…We choose: Spectral Filtering (SPF) [8], Bayes-Estimated Data Reconstruction (BE-DR) [9] and Multiple Miner attack with Fusion (MDMF) [10]. We use the SPF method because it has a good track record in reconstructing original data based on additive perturbation; it is based on eigenvalues of a covariance matrix and the theory of random matrices [8].…”

“…ii. We evaluate the success of attack by calculating , , for all 1≤ w, j ≤ √m ( [10]) and find that 5337 elements out of 6400 elements of the matrices have satisfied the inequalities. In this case, the attacker has obtained 83.39% of the original data and failed to recover 16.61 % of it.…”

“…We choose BE-DR for its ease of calculation and also because of its similarity to the calculations of SPF. The MDMF method is a combination of multiple data mining [10] and fusion techniques [10]; we use WEKA software [11] for data mining techniques in this method.…”