A new document classification algorithm against malicious data leakage attacks

Abstract

Nowadays it is important to store sensitive data and restrict its usage only to authorized people or institutions. In general, solutions for Data Leakage Prevention (DLP) ignores malicious attacks on documents and algorithms using fingerprinting and regular expressions are used. However, content-based attacks are successful evading those algorithms. In this paper an algorithm robust against malicious content-based attacks is proposed, which is independent of the attack executed. Transposition, sentence structure alteration, modification, obfuscation attacks are taken into consideration within the scope of paper. N-gram, charactergram, k-skip-n-gram and LSA methods are used in the feature extraction step, for having better classification results under attacks. The extracted features are passed to a Vote Classifier consisting of Support Vector Machine, Random Forest and Multi-Layer Perceptron classifiers. Additionally, the effects of instrumenting Spell-Correction in different steps of the algorithm is evaluated, which is effective against modification attacks.