Análisis y diseño de un modelo predictivo para detección de phishing basado en url y corpus del correo electrónico.

Abstract

One of the most reported cybercrimes worldwide is phishing, for this reason, various anti-phishing systems (APS) are currently being developed to identify this attack in online communication systems. Despite the best efforts, this attack continues unabated, having as causes, the erroneous detection in the zero-day attack, the high computational cost and the high rates of forgery. Although the Machine Learning (ML) approach has achieved a favorable accuracy rate, the choice and performance of the feature vector should be considered a key point to obtain a high level of accuracy. In this work, we propose a predictive model based on ML and analyze the efficiency of some anti-phishing schemes that served to understand this issue. The proposed model consists of a feature selection module that is used for the construction of the final vector. These characteristics are extracted from the URL, the web page properties and the email corpus, using a system based on incremental components to present the resulting vector. The system uses classification models, Random Forest and Naïve Bayes, which have been trained on the vector of traits. The experiments were based on datasets comprised of phishing and benign instances. Using cross-validation, the experimental results indicate a precision of 97.5% for the bases mentioned in other works, while for the approach of this research at the local level a precision of 96.5% was obtained.