Parameter-Efficient Hybrid Architectures for Multimodal Phishing Detection: Insights from the MTLP Dataset

This paper presents a comprehensive evaluation of phishing detection architectures, systematically transitioning from single-modality text classifiers to parameter-efficient hybrid multimodal systems. Using the MTLP dataset, a large-scale collection containing 100,000 legitimate and phishing webpage samples, we systematically evaluated multiple model configurations. The dataset includes Uniform Resource Locators (URLs), HyperText Markup Language (HTML) content, and WHOIS metadata. Each experiment targeted a specific research question, examining the impact of Parameter-Efficient Fine-Tuning (PEFT) using Low Rank Adaptation (LoRA), feature fusion strategies, and feature engineering on detection accuracy and efficiency. The optimal configuration (V5.2) integrated a language-specific Turkish BERT encoder with 14 domain-specific URL features, achieving an F1-score of 99.14% and outperforming text-only, image-based, and WHOIS-based baselines. Our experiments show that increasing architectural complexity (e.g., cross-attention) does not necessarily lead to better performance; instead, a balanced fusion of linguistic and structural cues yields optimal results. We also observed a clear trade-off between parameter count and accuracy, as lightweight models with only 1,538 trainable parameters achieved near state-of-the-art performance. Error analysis revealed that even advanced systems struggle with structurally ambiguous or content-sparse webpages. Overall, the findings highlight the potential of domain-adapted hybrid architectures as a robust, efficient, and explainable approach to phishing detection.