Scholarlink

Phishing email detection is a crucial aspect of cybersecurity, designed to defend people and institutions from unwanted attacks that try to steal confidential data. Traditional detection methods, such as rule-based systems and statistical analysis, often struggle to identify sophisticated and evolving phishing attempts. In this study, we explore the use of Large Language Models (LLMs) for phishing email detection, focusing on the analysis of email header features. LLMs, pre-trained on vast amounts of textual data, are adept at recognizing subtle patterns and linguistic nuances that are often indicative of phishing attacks. Our approach involves fine-tuning a state-of-the-art LLM to classify phishing emails based on their header features. We use the well-known Enron Email Dataset and Nazario Phishing Corpus. In the experiment, we found that LLMs outperformed conventional methods, achieving the highest accuracy, precision, recall, and F1-scores of 99.81%, 99.66%, 99.77%, and 99.71%, respectively. According to the findings, this model delivers the best performance among all currently available models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Phishing Email Detection using Header Features Leveraging Large Language Models (LLMs)

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chanchal Patra,
  • Debasis Giri,
  • Sutanu Nandi,
  • Tanmoy Maitra

摘要

Phishing email detection is a crucial aspect of cybersecurity, designed to defend people and institutions from unwanted attacks that try to steal confidential data. Traditional detection methods, such as rule-based systems and statistical analysis, often struggle to identify sophisticated and evolving phishing attempts. In this study, we explore the use of Large Language Models (LLMs) for phishing email detection, focusing on the analysis of email header features. LLMs, pre-trained on vast amounts of textual data, are adept at recognizing subtle patterns and linguistic nuances that are often indicative of phishing attacks. Our approach involves fine-tuning a state-of-the-art LLM to classify phishing emails based on their header features. We use the well-known Enron Email Dataset and Nazario Phishing Corpus. In the experiment, we found that LLMs outperformed conventional methods, achieving the highest accuracy, precision, recall, and F1-scores of 99.81%, 99.66%, 99.77%, and 99.71%, respectively. According to the findings, this model delivers the best performance among all currently available models.