Semantic Similarity Detection For Data Leak Prevention

Abstract

To counter data breaches, we introduce a new data leak prevention (DLP) approach. Unlike regular expression methods, our approach extracts a small number of critical semantic features and requires a small training set. Existing tools concentrate mostly on data format where most defense and industry applications would be better served by monitoring the semantics of information in the enterprise. We demonstrate our approach by comparing its performance with other state-of-the-art methods, such as latent dirichlet allocation (LDA) and support vector machine (SVM). The experiment results suggest that the proposed approach have superior accuracy in terms of detection rate and false-positive (FP) rate.