On persuasion in spam email: A multi-granularity text analysis

Electronic mail (email) is one of the most popular communication media for direct and private communication. Being typically a free service and anonymity-friendly, massive spam email campaigns are common. Nowadays, spam email encompasses scam, phishing, malware distribution, and various other cybersecurity threats. Within these emails, recipients frequently encounter social engineering techniques aimed at persuading them to take an action, such as clicking on a hyperlink, opening an attachment or responding. In this paper, we conduct a study on supervised models to identify persuasion (binary classification) and to identify the specific persuasion techniques that are commonly used in spam email (multilabel classification). To achieve this, we develop systems capable of spotting persuasion in spam emails based on natural language processing techniques. We approach this challenging task at different levels of granularity: full email, sentences and specific text snippets (i.e. text fragments composed by one or more words, typically shorter than a sentence). We replicate and adapt two methodologies used to detect propaganda in news articles. Additionally, we build a custom spam email dataset, and fine-tune pre-trained RoBERTa-based transformer models to tackle the sentence level detection. This allows us to determine how extensively spam emails rely on persuasion to achieve their goals and, if so, to identify those techniques that would be employed for user protection and cybersecurity improvements.