Abstract

To address the growing threat of phishing attacks in e-commerce, this study proposes a detection model based on the Performer architecture and designs a two-stage feature selection strategy to reduce feature redundancy and improve model generalization. This paper proposes a two-stage feature selection strategy that combines Pearson correlation coefficients with importance ranking based on random forest to filter the original features. Subsequently, highly correlated features with low importance are removed, resulting in a reduced feature subset. Next, this paper constructs the Performer-based detection model, which is then trained on a public dataset. To improve generalization and prevent overfitting, the model incorporates dynamic learning rate adjustment and early stopping during training. Experimental results show that the model achieves a recall of 0.9724 for phishing samples with a false negative rate of 2.8% on the test set, confirming its effectiveness in real-world scenarios. Overall, this study demonstrates the effectiveness of the Performer model in detecting e-commerce phishing attacks, providing a practical solution for enhancing platform security and strengthening supply chain resilience, while also opening new avenues for applying deep learning methods in the field of cybersecurity.

Download

Share

COinS