隨著社交工程與釣魚技術演進,具高度語意偽裝的中文動態釣魚郵件已難以透過傳 統靜態關鍵字機制有效攔截。為此,本研究提出一套創新的中文釣魚郵件內容分析框架,結合檢索增強生成(RAG)與動態微調機制,並以內政部警政署165 反詐騙網站之即時情資與在地化資安個案作為動態核心知識庫,旨在克服現行大型語言模型(LLMs)面對新興威脅時的幻覺現象。本框架具有二大特色:第一,透過RAG動態擷取即時情資,賦予地端模型辨識最新手法與在地化時事誘餌的能力;第二,利用動態微調優化 Llama3-8B 與Mistral-7B 等開源模型,強化對中文社交工程語意屬性與行為模式的理解。本研究針對中文變種釣魚郵件防禦盲點,提出動態融入即時威脅情資之地端大型語言模型(LLM)防禦框架。實驗採三階段消融分析(Ablation Study),並建立涵蓋傳統釣魚郵件與臺灣本土真實詐騙案例共8,000筆之獨立測試資料集,系統性評估不同提示策略、推論方式及知識檢索組合之效能。實驗結果顯示,以Few-shot Prompt 、 Chain-of-Thought(CoT)推論及165 反詐騙情資結合歷史訓練集之混合RAG檢索策略為最佳配置,相較於傳統機器學習分類器及Zero-shot基準模型(F1 = 53.09%),本研究方法可達94.94%的準確率、94.99%的F1分數,並將誤報率控制於6% 。研究結果證實,將即時威脅情資動態整合至地端大型語言模型,可有效提升中文變種釣魚郵件之偵測效能,並強化企業與組織面對新型社交工程攻擊之主動防禦能力。;This study proposes a Chinese phishing email detection framework that combines Retrieval-Augmented Generation (RAG) and dynamic fine-tuning. The framework uses real time threat intelligence from Taiwan’s National Police Agency 165 Anti-Fraud Portal and localized cybersecurity cases to support locally deployed LLMs, including Llama-3-8B and Mistral-7B. RAG retrieves updated threat information, while dynamic fine-tuning improves the models’ understanding of Chinese social engineering semantics and behavioral patterns. A three-stage ablation study was conducted on 8,000 phishing email samples from conventional phishing emails and Taiwan-based fraud cases. Results show that the best configuration, combining Few-shot prompting, Chain-of-Thought reasoning, and hybrid RAG retrieval, achieved 94.94% accuracy, 94.99% F1-score, and a 6% false positive rate,outperforming a Zero-shot baseline (F1=53.09%). The findings indicate that integrating real-time threat intelligence into local LLMs can improve Chinese phishing email detection and support organizational defense against evolving social engineering attacks.