This study investigates the effectiveness of attention-based tabular deep learning for detecting malicious PDF files in phishing-driven cyberattacks. Using PDF file samples from the CIC-Trap4Phish-2025 dataset, the performance of state-of-the-art multi-layer perceptron (MLP) models is benchmarked against a TabTransformer architecture across multiple data scales ranging from 5k to over 19k samples. The experimental results demonstrate feature reduction of up to 70% from the initial top-10 features without performance degradation. While the MLP baseline achieves competitive performance in low-data regimes, its learning saturates and slows as the dataset size increases. In contrast, the TabTrans former exhibits faster convergence and superior scalability. Overall, compared to the MLP baseline, the TabTransformer improves accuracy by up to 12.94% (97.7% vs. 86.5%), F1-score by 13.24%, and ROC-AUC by 4.54% in low-data settings. Evaluation across varying data scales confirms that attention-enabled tabular architectures provide a robust and efficient solution for document-level malware classification in modern cybersecurity systems.
Ayodeji et al. (Wed,) studied this question.