A dataset for web page classification

• Main Authors: Asoy, Julian Y., Domingo, Kenneth Vincent G.
• Format: text
• Language: English
• Published: Animo Repository 2016
• Subjects: Web sites > Security measures; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/14967
• Institution: De La Salle University

The internet is becoming larger every day, giving people public access to millions of web pages. Unfortunately, in this vast domain of web pages, the presence of malicious web pages is inevitable. Thus, the detection of these malicious web pages can contribute to the security and protection of internet users. There have been various studies regarding the malicious web page detection. In these studies, a variety of significant web page features are extracted from collected URLs and web pages to characterize malicious and benign web pages. In addition, different machine learning techniques were used together with the features to detect malicious web pages.;"This study focuses on the data collection and feature extraction processes involved in malicious web page detection by creating a usable dataset of URLs, web page contents, and features. The dataset contains 58,654 usable web pages. Of which , 61% are benign web pages and 39% are malicious. Each web page has 24 features. 15 of these features have been used in previous studies, while the other 9 are experimental features based on HTML5. The dataset was used in 3 tests, each with different sets of features in order to evaluate their effectiveness. A total of 6 types of machine learning techniques, classified as parametric (naive bayes (NB)) and logistic regression (LR) and non-parametric (decision tree (DT)), support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN)), are used to experiment and test the usefulness of the datasets. Results show that non-parametric classifiers performed better in terms of kappa and accuracy. However, with an unbalanced dataset, classifiers are evaluated based on their precision and recall. Overall results showed that the SVM yielded the highest precision and recall values, with 81.92% and 73.4% respectively.