Research On Phishing Website Detection Based On Neural Network And Attention Mechanism

With the commercialization of 5G networks,people have begun to rely more and more on the network.A large quantity of information is generated within the network daily.These data include some personal information of network users,bank card numbers,payment passwords and so on.The security data interaction between users and the website becomes particularly important.At present,phishing websites became a serious threat to network security because of their short survival time and nice damage.Phishing uses social engineering techniques like email and SMS to steal personal data from users by disguising phishing URLs as URLs of legitimate websites.Therefore,it is particularly important to establish a fast and effective phishing network detection model.The current detection methods are mainly based on machine learning and deep learning.The difference between the two methods lies in the difference in feature engineering.Feature engineering based on machine learning methods generally rely on prior knowledge,while methods based on deep learning use complex neural networks to automatically extract features.According to the comparison of multiple models,the accuracy of the neural network-based method is higher.One problem with both methods is that there will be negative features in these features.These features not only cannot effectively distinguish phishing websites,but may also reduce the detection accuracy of the model.Based on the above analysis,this dissertation proposes a lightweight phishing detection method CCBLA(char CNN and Bi-LSTM with attention mechanism),which extracts features based on a deep learning algorithm and uses an attention mechanism for feature selection.In the CCBLA model,deep learning methods are used to automatically extract features from the target URL.These features have different degrees of importance in the phishing detection process,and then the attention mechanism is used to perform feature selection operations to remove those features that do not improve detection accuracy.Two data sets of different scales are used to test the performance of the proposed CCBLA model.From the experimental results,it can be seen that this methodology is correct in sleuthing phishing attacks and has sensible detection potency.The main work of this dissertation is as follows:(1)CNN and Bi-LSTM neural network are used to automatically extract local features and contextual semantic structure features of URLs.It will effectively forestall criminals from by choice planning URLs as a result of they grasp the options extracted by hand-supported machine learning ways.additionally,these options don't rely on third-party services,which might greatly improve the detection speed.Since the phishing site detection model does not require that high computational resources,the CCBLA model is also suitable for deployment on mobile platforms where computational resources are relatively scarce.(2)An attention score calculation method based on the TF-IDF algorithm is proposed.Since the feature dimension extracted by deep learning algorithms can be very large and not all features are positive features this requires feature selection.So we propose the TF-IDF attention score by combining the TF-IDF algorithm to improve the attention score calculation of the traditional attention mechanism.The traditional attention score calculation process needs to initialize a parameter as it happens and correct it in the continuous iteration of the neural network.Our method first calculates the attention score and does not participate in the training,which reduces the training parameters of the neural network.Feature selection by improving the attention mechanism not only discards some negative features,but also reduces the complexity of the model,and the speed and accuracy of model detection have been greatly improved.