Phishing emails pose a significant threat to cybersecurity worldwide. There are already tools that mitigate the impact of these emails by filtering them, but these tools are only as reliable as their ability to detect new formats and techniques for creating phishing emails. In this paper, we investigated how traditional models and transformer models work on the classification task of identifying if an email is phishing or not. We realized that transformer models, in particular distilBERT, BERT, and roBERTa, had a significantly higher performance compared to traditional models like Logistic Regression, Random Forest, Support Vector Machine, and Naive Bayes. The process consisted of using a large and robust dataset of emails and applying preprocessing and optimization techniques to maximize the best result possible. roBERTa showed an outstanding capacity to identify phishing emails by achieving a maximum accuracy of 0.9943. Even though they were still successful, traditional models performed marginally worse; SVM performed the best, with an accuracy of 0.9876. The results emphasize the value of sophisticated text-processing methods and the potential of transformer models to improve email security by thwarting phishing attempts.
The automated detection of deceptive language is a crucial challenge in computational linguistics. This study provides a rigorous comparative analysis of three tiers of machine learning models for detecting instructed deception: traditional machine learning (SVM), fine-tuned discriminative models (BERT), and in-context learning with generalist Large Language Models (LLMs). Using the “cross-cultural deception detection” dataset, our findings reveal a clear performance hierarchy. While SVM performance is inconsistent, fine-tuned BERT models achieve substantially superior accuracy. Notably, a multilingual BERT model improves cross-topic accuracy on Spanish text to 90.14%, a gain of over 22 percentage points from its monolingual counterpart (67.20%). In contrast, modern LLMs perform poorly in zero-shot settings and fail to surpass the SVM baseline even with few-shot prompting, underscoring the effectiveness of task-specific fine-tuning. By transparently addressing the limitations of the solicited, low-stakes deception dataset, we establish a robust methodological baseline that clarifies the strengths of different modeling paradigms and informs future research into more complex, real-world deception phenomena.