Figure 1:
Architecture of hybrid generative-discriminative classifier with SE-GTE. SE-GTE, self-evolving generative threat engine.
Table 1:
Example transformation of a URL through the proposed feature extraction stages
| Stage | Sample output |
|---|---|
| Raw URL | ‘https[//]www[.]amazon[.]com/gp/help/customer/display.html’ |
| Character Tokenization | <STA> h t p s [ / / ] w [ . ] a m a z o n … h t m l <SEP> |
| Character Encoding | [2,11,23,23,19,22,42,40,40,43,27,27,27,…,28,3] |
| Generative Encoder Output | [0.41, −0.27, 1.13, 0.08, −0.66, …, 0.09] |
| Variational Encoder (Mean) | [0.38, −0.31, 1.05, 0.12, −0.60, …, 0.11] |
| Variational Encoder (Variance) | [0.04, 0.06, 0.02, 0.05, …, 0.04] |
Table 2:
Comparative performance of the proposed VGA-HyPhishNet and baseline models
| Model | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|
| VGA-HyPhishNet | 0.9873 | 0.9300 | 0.9497 | 0.9397 |
| URL-Transformer | 0.9556 | 0.7803 | 0.7999 | 0.7899 |
| BiLSTM-URL | 0.9444 | 0.7100 | 0.7896 | 0.7477 |
| CharCNN-URL | 0.9387 | 0.6803 | 0.7799 | 0.7267 |
Figure 2:
Confusion matrix.
Figure 3:
ROC curve. FPR, false positive rate.
Figure 4:
Inference time comparison across phishing detection models.
Figure 5:
Training cost comparison across phishing detection models.
Table 3:
Ablation study and parameter sensitivity of proposed study
| Setting/variant | Key change | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|---|
| Full Model (VGA-HyPhishNet) | Encoder + Variational Modeling + SE-GTE + Hybrid Fusion | 0.9873 | 0.9300 | 0.9497 | 0.9397 |
| w/o VGM (No variational modeling) | Variational modeling removed; deterministic features used | 0.9728 | 0.8850 | 0.9040 | 0.8944 |
| w/o SE-GTE | Synthetic threat generation removed | 0.9762 | 0.9021 | 0.9215 | 0.9117 |
| w/o Hybrid Fusion | Only discriminative classification used | 0.9709 | 0.8724 | 0.9280 | 0.8993 |
| Generative-only | Only generative decision used | 0.9584 | 0.8202 | 0.9020 | 0.8591 |
| Discriminative-only | Only classification-based decision used | 0.9687 | 0.8658 | 0.9149 | 0.8897 |