SOCIALBERT a Transformer based Model Used for Detection of Social Engineering Characteristics

Security threats have escalated in recent years based on social engineering attacks. In these attacks, bad actors target organizations, businesses, and private individuals by luring them into clicking on a link either in a text message, email message, or a document. Social Engineering is the act of using psychological tricks to gain the trust of humans leading to the release of sensitive information which can be used for various nefarious activities by the bad actors. These types of attacks have proven to be the genesis of most of the malicious attacks launched against both physical and virtual IT systems by tricking humans into allowing these attacks. It is scarce to find widely established machine learning techniques specifically designed for identifying social engineering features in messages. To help tackle this problem, this research paper aims to explore the detection of social engineering tactics in text messages by fine-tuning a pre-trained Machine Learning Transformer model, BERT to produce our model called SOCIALBERT. SOCIALBERT (Social Engineering Detection Bert) achieves a 97.55 percent accuracy as compared to DistilBERT, Bert-Base multilingual, and Naive Bayes Models which achieved 55.40 percent, 60.23 percent, and 95.58 percent accuracies respectfully.