A Lightweight Approach for Phishing Email Detection in Turkish with LLMs

Phishing attacks are one of the most common information theft techniques, and they are a significant threat to all users. An email, SMS and websites can be used to access the user’s confidential data through misleading information, imitated websites, and links. Since one of the most effective security approaches is prevention, various methods are developed such as implementing large language models (LLMs) for the detection of these attacks. Integration of these models into different languages is required to increase the usability and benefits. In this research, a new Turkish phishing email dataset is created via translation of an existing dataset. Then, the ELECTRA Small Turkish LLM model is chosen to be trained after the evaluation of different models. Low-Rank Adaptation (LoRA) fine-tuning method is used to increase the performance. Although the model obtains accuracy 97% and an F1 score 97% with the English dataset, it performed worse on the Turkish data set. Its performance increased to 92.8% accuracy and 93.1% F1 score after training with LoRA. Then, using the Turkish pre-trained ELECTRA Small Turkish model, better results are achieved, which are 98.3% accuracy and 98.4% F1 score. These results highlight that fine-tuned LLMs or transfer learning approaches can be efficiently used for phishing email detection in different languages.