Social Engineering Attack Detection Using Natural Language Processing and Machine Learning Techniques

Detection of social engineering attacks, including phishing, pre-texting, baiting, and tailgating, is challenging. Our approach is based on machine learning techniques to identify these threats effectively. First, we collect a data set that is sourced from the Text Retrieval Conference (TREC) corpus, where each entry is provided as plain text. Then, we perform data cleaning to remove irrelevant elements such as HTML tags, special characters, and stop words. Following this, we do data pre-processing that converts categorical features into numerical features. The pre-processed data is divided into two parts, namely, training and testing subsets. Then, we train three models, namely, the Naive Bayes Classifier, Support Vector Machine (SVM), and Random Forest on the training subset. These models are evaluated for the testing subset using metrics that include accuracy, precision, recall, and F1 score. From the evaluation, we have found that the SVM consistently outperforms the other models with accuracy (98 %), precision (98.50%), recall (98.70%), and F1-score (98.60%).