BERT Against Social Engineering Attack: Phishing Text Detection

Social engineering attack uses a wide range of human interaction tricks with the goal of achieving sensitive information. Certain tricks involve sending malicious SMS to the victim where the victim is convinced by the SMS, making a security mistake by clicking a malicious link or giving away confidential information. Machine learning algorithms for spam filtering are effective measures against such SMS spam. This paper demonstrated a novel universal spam detection model using pre-trained Google bidirectional encoder representations from Transformers (BERT) for classifying spam SMS in real-time scenarios. Subsequently, different classification techniques on the datasets are evaluated, based on their accuracy, precision, and recall. An overall accuracy reached 99%, with an F1 score of 0.97. The results and implications confirmed that the distilled BERT is an effective approach for spam detection.