Utilizing Convolutional Neural Networks and Word Embeddings for Early-Stage Recognition of Persuasion in Chat-Based Social Engineering Attacks

Social engineering is widely recognized as the key to successful cyber-attacks. Chat-based social engineering (CSE) attacks are attracting increasing attention because of recent changes in the digital work environment. Sophisticated CSE attacks target human personality traits, and persuasion is regarded as the catalyst to successful CSE attacks. To date, research in social engineering has mostly focused on phishing attacks, neglecting the importance of chat-based software. This paper describes the design and implementation of a persuasion classifier that utilizes machine learning and natural language processing techniques. For this purpose, a convolutional neural network was trained on a chat-based social engineering corpus (CSE Corpus), specifically annotated for recognizing Cialdini's persuasion principles. The proposed persuasion classifier network, named CSE-PUC, can determine whether a sentence carries a persuasive payload by producing a probability distribution over the sentence classes as a persuasion container. The present study is expected to contribute to our understanding of utilizing existing machine learning models and integrating context-aware information into real-life cyber security threats. The experimental application results reported in this work confirm that the approach taken can recognize persuasion methods and is thus able to protect an interlocutor from being victimized.