Achieving probabilistic anonymity against one-to-multiple linkage attacks

Randomization methods widely applied for privacy-preserving data mining are generally subject to reconstruction attack, linkage attack, and semantic-related attacks. A probabilistic anonymity definition has been proposed in [1] to defend against the linkage attack in which the attacker links the same randomized record to all of the original records. In this paper we name this type of attack as Multiple (original records) to One (randomized record) attack, while focus on another attack that has not been researched before, i.e. One (original record) to Multiple (randomized records) attack. The latter is different from the former in that it does not require the attacker to know the distribution and all values of quasi-identifiers in original records, and thus is easier to be launched by the attacker. To defend against this attack we propose a novel probabilistic anonymity concept different from [1]. We achieve this anonymity goal on a hybrid model combining random projection and random noise addition. We also analyze the security properties of this model against the other common types of attacks. Compared with existing work in randomization, k-anonymity and differential privacy, our work achieves the holistic aim of higher security, higher efficiency and higher data utility, and demonstrates very promising applications in large-scale and high-dimensional data mining in clouds.