SpamLLM: Leveraging Large Language Models for Robust Spam Email Classification

Spam email detection remains an ongoing challenge due to the increasing sophistication and evolving tactics employed by spammers. Traditional rule–based and machine learning (ML) detection methods have demonstrated limitations in adaptability and generalization. This paper proposes SpamLLM, a multimodal spam detection framework that leverages a frozen large language model (LLM) backbone integrated with semantic embeddings and structured metadata extracted from email headers and body statistics. Rigorous evaluations are conducted across four widely used spam detection datasets, and SpamLLM is compared against classical ML, deep learning, and pretrained Transformer‐based baseline models. The experimental results demonstrate that SpamLLM outperforms existing methods and achieves state‐of‐the‐art performance in accuracy, precision, recall, and F1 score. Notably, SpamLLM excels in handling diverse spam content and provides robust detection across various datasets. These findings underscore the potential of multimodal fusion approaches for advancing spam classification systems and suggest promising directions for future research in email security.