Fraud Detection in Web Content Using Machine Learning and Natural Language Processing

The rise of Large Language Models offers new opportunities for online frauds but also new techniques for detecting them. This paper analyses if the classical approaches still work on novel web frauds and what is the improvement offered by the semantic approaches. A new dataset was build, crawling recent fraudulent and benign web pages. By taking into account features extracted from both the page’s URL and content, the semantic approach using BERT obtained the best accuracy of $98.1 \%$, but was also the slowest, compared to other models that achieved slightly lesser scores but are more than 1000 times faster.