以字元細化分割與BPNN辨識檢測CAPTCHA防護機制有效性之研究

Abstract

由於電子商務之興起，交易網站若遭受入侵將產生巨額損失，因此需要有效的身份識別與驗證機制，以確保交易的安全性及不可否認性。傳統廣泛採用帳號、密碼的驗證機制做為身份識別之用����然而，密碼的驗證機制很容易遭受程式以����典法或暴力法進行攻擊。為了避免資訊系統遭受惡意程式的攻擊，因此����展出以CAPTCHA (Completely Automated Public Turing Tests to Tell Computers and Humans Apart)的人機辨識輔助驗證機制，以防止惡意程式直接輕易對密碼進行猜測攻擊。CAPTCHA是一種由電腦����辨互動對象是電腦還是人的驗證機制，用以阻斷全自動機器人程式的惡意攻擊行為。本研究將針對大型的電子商務平台進行CAPTCHA的人工智慧辨識，進行其有效性之檢測。本研究透過字元筆劃軌跡細化處理，強化字元的分割作業，將可分割的字元擷取出來，再以倒傳遞類神經網路(Back Propagation Neural Network, BPNN)作為字元辨識的工具。本研究除透過完整字元圖像進行訓練外，並增加擷取圖像90%與80%作為輔助網路之訓練樣本，希望透過強化學習字元部分特徵，整合三個類神經網路，以增加辨識的準確率。根據本研究實驗結果，應用單一網路的字元辨識率為94%，若整合三個BPNN網路的字元辨識率則可提升至96%，驗證了透過截取不同比例的特徵圖形，進行多網路辨識決策，將有助於提升單一BPNN網路的辨識率。也顯示現有應用於電子商務的CAPTCHA機制需持續檢測及強化，以發揮充足之保護效果。The Internet developed rapidly along with the rise of e-commerce. In the past, Password-based authentication schemes are the most widely used mechanisms over e-commerce. Password-based authentication schemes may be weak to the dictionary attack. CAPTCHA authentication schemes are proposed for prevent automated script attacking. Past researchers have never proposed methods to solve overlapping and connecting problems. Therefore, we want to propose an effective method to solve overlapping and connecting problems. We detect the CAPTCHA and extract the divisible parts from the CPATCHA firstly. Then we use thinning method to increase the gap of characters. After extracting characters, we recognize the characters by using back propagation neural network (BPNN). In order to improve the problem of false positives, we will train two additional networks. I trained the network for the normalized image firstly. Then we capture the normalized image 90% and 80% for the other networks' training samples. We hope that the other two networks will learn same features of the part of characters. In this way, there are some bases for adjustment when single network answer wrong. Finally, we integrated the tree networks by voting. According to the experiment, the recognition rate of single network is 94% and the recognition rate of integrated three networks is 96%. It prove that multi-networks are better than single network.