Privacy-Preserving Data Mining in Presence of Covert Adversaries

Disclosure of the original data sets is not acceptable due to privacy concerns in many distributed data mining settings. To address such concerns, privacy-preserving data mining has been an active research area in recent years. All the recent works on privacy-preserving data mining have considered either semi-honest or malicious adversarial models, whereby an adversary is assumed to follow or arbitrarily deviate from the protocol, respectively. While semi-honest model provides weak security requiring small amount of computation and malicious model provides strong security requiring expensive computations like Non-Interactive Zero Knowledge proofs, we envisage the need for 'covert' adversarial model that performs in between the semi-honest and malicious models, both in terms of security guarantee and computational cost. In this paper, for the first time in data-mining area, we build efficient and secure dot product and set-intersection protocols in covert adversarial model. We use homomorphic property of Paillier encryption scheme and two-party computation of Aumann et al. to construct our protocols. Furthermore, our protocols are secure in Universal Composability framework.