A Hybrid Learning of Lexical and Language Processing for Domain Credibility Classification

Malicious domains and websites pose a significant threat to normal users and their increasing prevalence demands for early detection methods. More and more domains registered with malicious intent are becoming more excessively difficult to prevent and detect. Leveraging the recent powerful BERT based-language representation and conventional lexical feature representation, we introduce a novel hybrid learning model that utilizes both lexical characteristics and semantic language features of inspected domains for domain credibility classification. The proposed model employs a combination of lexical and language encoders to process lexical features like length, special character count, domain type, domain entropy, and domain digits while fine-tuning the pre-trained language models such as Vietnamese PhoBERT and multilingual XLM-RoBERTa to capture semantic information from the domain. Through the experimental results, the hybrid learning models outperform the baselines such as using solely lexical encoder or language encoder in differentiatioz between High or Low credibility domains.