Phishing attacks remain among the most prevalent cybersecurity threats because malicious URLs can be rapidly created and modified to impersonate legitimate services and deceive users. Although machine learning and deep learning approaches have achieved strong performance in phishing URL detection, several challenges remain, including generalization to previously unseen phishing samples, resilience to adversarial URL manipulation, effective feature representation, and the selection of decision thresholds that balance missed attacks against false alerts.
This thesis presents a Hybrid VAE-Attention framework for strong zero-day phishing URL detection under adversarial evaluation. The framework takes a number of different URL representations as input (character-level and word-level TF-IDF and Word2Vec embeddings, structural URL features, and other preprocessed numeric features), performs weighted feature fusion and hybrid feature selection to produce a compact 183-dimensional representation. The model combines reconstruction-only information, a stochastic latent bottleneck, a Q/K/V-based attention projection, and an auxiliary classification branch. The final decision is made via validation and weighted-score fusion, with the fusion and decision threshold jointly optimizing the F2-scores to achieve high phishing accuracy at the expense of some control over the false-positive rate. The proposed model was evaluated on the PhiUSIIL Phishing URL Dataset, containing 235,795 URLs, of which 100,945 are benign and 134,850 are phishing samples. The proposed model was fitted using benign training samples, while phishing samples were introduced during validation and testing. Its performance was compared with Isolation Forest, K-Means, and Random Forest. Robustness was further assessed using an augmented test set containing 10,000 simulated URL modifications generated through six lexical and structural manipulation techniques.
On the augmented test set, the proposed Hybrid VAE-Attention model achieved 99.15% Accuracy, 98.95% Precision, 99.93% Recall, an F1-score of 0.9944, and a ROC-AUC of 0.9991. It correctly identified 77,371 of the phishing-labeled samples with 54 false negatives and 820 false positives. The proposed model achieved the highest Accuracy, recall, and F1-score among all the models considered, whilst Random Forest achieved the highest Precision, the highest ROC-AUC, and the fewest false positives. A paired bootstrap significance test (1,000 resamples) demonstrated that the F1-score improvements of the proposed framework over Random Forest, K-Means and Isolation Forest were all statistically significant (p < 0.001 in all cases).
The technique-level adversarial assessment was carried out using the calibrated Hybrid VAE-Attention model on 500 phishing source URLs for each of the six modified techniques (Homoglyph Substitution, Typosquatting, Path Obfuscation, Parameter Injection, Subdomain Abuse and URL Encoding). The model achieved a 100.0% Detection Rate and 0.0% Evasion Rate across all six techniques.
The overall results show that the hybrid URL representation, combined with reconstruction-based anomaly information, stochastic latent learning, attention projection, and calibrated score fusion, achieves good performance for phishing URL detection under the experimental conditions. This should be regarded as evidence of robustness to previously unseen samples and to the simulated adversarial variations on the dataset rather than as evidence of real-world zero-day generalization under a strict temporal or campaign-held-out evaluation.
