A dual-phase deep learning framework for advanced phishing detection using the novel OptSHQCNN approach

Phishing attacks are now regarded as one of the most prevalent cyberattacks that often compromise the security of different communication and internet networks. Phishing websites are created with the goal of generating cyber threats in order to ascertain the user's financial information. Fake websites are frequently created and circulated online, which results in the loss of essential user assets. Phishing websites can result in monetary loss, intellectual property theft, damage to one's reputation, and disruption of regular business activities. Over the past decade, a number of anti-phishing tactics have been proposed to detect and reduce these attempts. They are still imprecise and ineffective, though. Deep Learning (DL), which can precisely learn the intrinsic features of the websites and recognize phishing websites, is one of the innovative techniques utilized to solve this issue.

In this study, we proposed a novel OptSHQCNN phishing detection method. Pre-deployment and post-deployment are the two phases of the proposed methodology. The dataset undergoes preprocessing in the pre-deployment phase, which includes data balancing, and handling invalid features, irrelevant features, and missing values. The convolutional block attention module (CBAM) then extracts the main characteristics from web page code and linkages. The red kite optimization algorithm (RKOA) selects the significant key attributes in the third stage. The final phase involves classifying the data using the Shallow hybrid quantum-classical convolutional neural network (SHQCNN) model. To improve the effectiveness of the classification approach, the hyperparameters present in the SHQCNN model are fine-tuned using the shuffled shepherd optimization algorithm (SSOA).

In the post-deployment phase, the URL is encoded using Optimized Bidirectional Encoder Representations from Transformers (OptBERT), after which the features are extracted. The retrieved properties are fed into a trained classifier. Next, a prediction of "phishing" or "Legitimate" is produced by the classifier. With a maximum of above 99% accuracy, precision, recall, and F1-score, respectively, the investigation's findings showed that the suggested technique performed better than other popular phishing detection methods. The creation of a security plugin for clients, browsers, and other instant messaging applications that operate on network edges, PCs, smartphones, and other personal terminals can be aided by these findings.