Detection and Analysis of Phishing Emails Using Time-Series Models
Year: 2026
Type: conference-paper
Abstract: With the rapid evolution of cyber-attacks, phishing emails generated by large language models (LLMs) have become increasingly sophisticated and prevalent. Traditional detection methods often struggle to capture long-range semantic dependencies and are susceptible to noise in unstructured text. This paper evaluates the performance of three advanced time-series and sequential architectures, Transformer, BigBird, and Mamba, in detecting phishing emails. We constructed a comprehensive dataset of 6,000 samples, integrating malicious emails from the Dark Web, legitimate corporate communications, and AI-generated phishing content. Experimental results demonstrate that while the Mamba model achieves the highest overall F1-score of 97.4% alongside remarkable training efficiency with O(n) complexity, the BigBird architecture also exhibits a highly competitive and robust F1-score of 96.5% due to its specialized sparse attention mechanism in handling extended sequences. This study provides a systematic comparison of model robustness against text noise and long-range dependencies, offering a high-performance framework for modern cybersecurity defense.
Institution: Tunghai University
Language: English
Topic: Spam and Phishing Detection
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204205377