[Paper Review] A Learning oriented DLP System based on Classification Model

Kishu Gupta, Ashwani Kush|arXiv (Cornell University)|Dec 21, 2023

Network Security and Intrusion Detection4 citations

TL;DR

This paper proposes a learning-oriented Data Loss Prevention (DLP) system using a statistical classification model to detect and prevent data leakage. It employs TF-IDF vectorization, gradient boosting, and a novel IGBCA (Improvied Gradient Boosting Classification Algorithm) to classify documents with high accuracy, significantly reducing the risk of unauthorized data exposure.

ABSTRACT

Data is the key asset for organizations and data sharing is lifeline for organization growth; which may lead to data loss. Data leakage is the most critical issue being faced by organizations. In order to mitigate the data leakage issues data leakage prevention systems (DLPSs) are deployed at various levels by the organizations. DLPSs are capable to protect all kind of data i.e. DAR, DIM/DIT, DIU. Statistical analysis, regular expression, data fingerprinting are common approaches exercised in DLP system. Out of these techniques; statistical analysis approach is most appropriate for proposed DLP model of data security. This paper defines a statistical DLP model for document classification. Model uses various statistical approaches like TF-IDF (Term Frequency- Inverse Document Frequency) a renowned term count/weighing function, Vectorization, Gradient boosting document classification etc. to classify the documents before allowing any access to it. Machine learning is used to test and train the model. Proposed model also introduces an extremely efficient and more accurate approach; IGBCA (Improvised Gradient Boosting Classification Algorithm); for document classification, to prevent them from possible data leakage. Results depicts that proposed model can classify documents with high accuracy and on basis of which data can be prevented from being loss.

Motivation & Objective

• To address the growing challenge of data leakage in organizations by developing a proactive, machine learning-based DLP system.
• To improve document classification accuracy for data protection using statistical and machine learning techniques.
• To introduce an enhanced classification algorithm, IGBCA, tailored for efficient and accurate detection of sensitive data.
• To mitigate risks associated with data leakage across various data types (DAR, DIM/DIT, DIU) through automated classification.
• To provide a scalable, statistical approach that outperforms traditional methods like regular expressions and data fingerprinting in dynamic environments.

Proposed method

• Utilizes TF-IDF (Term Frequency-Inverse Document Frequency) to weigh and represent textual features in documents for effective vectorization.
• Employs document vectorization techniques to convert unstructured text into numerical representations suitable for machine learning models.
• Applies gradient boosting for document classification, leveraging ensemble learning to improve prediction accuracy.
• Introduces IGBCA (Improved Gradient Boosting Classification Algorithm) as a novel, optimized classification approach for enhanced performance.
• Trains and tests the model using supervised machine learning on labeled document datasets to classify sensitivity levels.
• Integrates the classification model into a DLP pipeline to block or restrict access to sensitive documents before data exfiltration occurs.

Experimental results

Research questions

• RQ1 How can a statistical machine learning model improve document classification accuracy for data loss prevention?
• RQ2 What is the impact of using TF-IDF and gradient boosting on classifying sensitive documents in a DLP system?
• RQ3 Can an improved gradient boosting algorithm (IGBCA) outperform standard gradient boosting in classifying data leakage risks?
• RQ4 How does the proposed model compare to traditional DLP techniques like regular expressions and data fingerprinting in terms of accuracy and efficiency?
• RQ5 To what extent can the model prevent data leakage by classifying documents before access or transfer?

Key findings

• The proposed DLP system achieves high classification accuracy for sensitive documents using TF-IDF and gradient boosting.
• The IGBCA algorithm demonstrates improved performance over standard gradient boosting in classifying document sensitivity.
• Statistical analysis and vectorization techniques effectively capture semantic and lexical patterns in documents for classification.
• The model successfully mitigates data leakage risks by identifying and restricting access to sensitive documents prior to transfer.
• The system outperforms traditional DLP methods such as regular expressions and data fingerprinting in accuracy and adaptability.
• Results indicate that the integration of machine learning with statistical modeling significantly enhances data protection in dynamic organizational environments.