Web2VecPhishingDataset: A Multi-Space Feature Dataset for Web Phishing Research
Description
The Web2VecPhishingDataset is a collection of four machine-learning-ready data artifacts designed for research on phishing website detection, dataset construction, and feature-space comparison. The artifacts were generated using Web2VecPhishingPipeline (https://github.com/edytafraszczak/Web2VecPhishingPipeline) (https://doi.org/10.1016/j.softx.2026.103074), which supports data ingestion, URL and domain normalization, binary label management, Web2Vec-based feature extraction, integration of compatible external records, and feature-space harmonization. The release contains four complementary artifacts: - Web2VecFullSpace contains 496,850 rows and 236 columns, including 233 candidate features and 183 numeric machine-learning-ready features. It covers URL lexical, DNS, GEO/network, HTML, HTTP, Open PageRank, SSL/TLS, and WHOIS-derived features. The artifact contains 301,151 legitimate and 195,699 phishing observations. - Web2VecUrlSpace contains 5,666,075 rows and 101 columns, including 98 candidate URL lexical features and 97 numeric machine-learning-ready features. It contains 999,993 legitimate and 4,666,082 phishing observations. - Web2VecUrlMergedSpace extends the URL-based representation with compatible records from additional datasets. It contains 6,134,211 rows and 101 columns, including the same 98 candidate features and 97 numeric machine-learning-ready features as Web2VecUrlSpace. It contains 1,265,082 legitimate and 4,869,129 phishing observations. - Web2VecCommonSpace is an identifier-free harmonized representation based on 18 shared numeric lexical features. It contains 6,397,495 rows, including 1,483,189 legitimate and 4,914,306 phishing observations. It consists of the 18-feature projection of all 6,134,211 Web2VecUrlMergedSpace rows and 263,284 records that are available only in the common schema. The final Web2VecUrlSpace and Web2VecUrlMergedSpace files contain unique normalized URL identifiers. The documented uniqueness key combines the lower-cased protocol,nnormalized complete host, and effective port. Credentials, paths, queries,mfragments, letter case, and a terminal host dot do not define a separate key. Label 0 represents a legitimate observation, and label 1 represents a phishing observation. Labels apply to individual observations and should not be interpreted as indicating that every path, subdomain, registered domain, hosting platform, or associated service is malicious. The four artifacts differ in scale, feature richness, identifier availability, and source composition. To reduce information leakage during model evaluation, all observations associated with the same normalized complete host should remain in one data partition. Source-held-out, temporally separated, and independently collected test sets are also recommended.
Files
Institutions
- Military University of Technology in WarsawMazovia, Warsaw