自從2020年全球受到Covid-19影響，民眾對非接觸式的數位交易需求增加，許多服務由實體轉至雲端，網路上的金融交易量大幅增加，居家上班的工作型態，也讓公司的機敏資料暴露在更容易受到攻擊的環境之下，網路釣魚在 2020 年開始呈現急速成長，至2021年底以成長兩倍。傳統的釣魚網站攻擊檢測模型依賴啟發式規則尋找特徵，結合機器學習進行預測，本研究想要提出一種模型來補足傳統模型的功能。 本研究提出一種使用長短期記憶神經網路(long short term memory，LSTM)和啟發式規則的混合特徵模組，該模組以啟發式規則收集重要特徵，再以LSTM萃取網址特徵彌補啟發式規則的弱點，最後使用類神經網路(Neural Network，NN)進行預測。實驗中發現LSTM確實能找出潛藏的特徵以輔助模型進行更準確的釣魚網站預測，準確度(ACC)為0.997。 ;Since the global impact of Covid-19 in 2020, the demand for contactless digital transactions has increased, many services have moved from the physical to the cloud, financial transactions on the Internet have increased dramatically, and the work-from-home work style has exposed companies′ sensitive data to a more vulnerable environment. The number of phishing attacks will begin to grow rapidly in 2020 and will increase twofold by the end of 2021. Traditional phishing attack detection models rely on heuristic rules to find features and combine with machine learning to make predictions. This study proposes a hybrid feature module using long short term memory (LSTM) and heuristic rules. The module collects important features with heuristic rules, and then uses LSTM to extract URL features to supplement the heuristic rules. Weaknesses, and finally use a neural network (Neural Network, NN) for prediction. In the experiment, it is found that LSTM can indeed find hidden features to assist the model to make more accurate predictions of phishing websites, with an accuracy (ACC) of 0.997.
