Lightweight privacy-Preserving data classification

Abstract

Internal attacks are of a huge concern, because they are usually delicately masqueraded under harmless-looking activities, which are very difficult to detect. Machine learning techniques have been successfully applied to identify insider threats. However, they may violate user privacy since they can legally access user’s sensitive information. To preserve user privacy, encryption algorithms have been lately exploited as a powerful tool, to hide private data in a multiple-party collaboration. A combination of encryption and data mining techniques raises high computational complexity. Hence, in order to improve the system’s performance while securing both user’s private data and the classifier, we propose a new secure data analysis protocol, namely SmartClass, by adopting the garbled circuit technique to speed-up the system performance. We developed an efficient encryption step that exploits the additive homomorphism and best properties of the binary Elliptic Curve Cryptography (ECC) algorithm, while keeping the protocol highly secure. We implemented the proposed system and study its effectiveness. Experimental results show the proposed approach is very promising.