关键词: 网络安全, 钓鱼邮件检测, 预训练语言模型, 低秩适配, 语义特征提取

Abstract: Phishing email detection is critical in cybersecurity, as it faces significant challenges due to the diverse and complex nature of phishing emails. This paper proposes a phishing email detection method integrating the pretrained language model DistilBERT with LowRank Adaptation (LoRA). DistilBERT is used to extract deep semantic features from email text, while LoRA finetunes a small number of parameters, thereby reducing the dependence on largescale labeled data and enhancing the model generalization. Experimental results show that compared to traditional machine learning methods and deep learning models (such as RNN, LSTM, and Bidirectional LSTM), DistilBERT+LoRA outperforms them in key metrics including accuracy, precision, recall, and F1score, achieving 96% accuracy and 97% F1score, which significantly surpassing comparative models. Additionally, it demonstrates better balance between precision and recall than other deep learning models, particularly demonstrating robustness and adaptability in detecting complex phishing emails. Experiments further reveal that the model’s performance improves with the increase in LoRA’s rank parameters. By leveraging the powerful feature extraction capabilities of pretrained language models and the efficient finetuning advantages of LoRA, this method provides an innovative and effective solution for accurate and efficient phishing email detection.

Key words: cybersecurity, phishing email detection, pretrained language model, LoRA(lowrank adaptation)