Cybercrime Prediction via Multi-Modal Learning from Behavioral, Network, and Textual Data

Abstract

     The increasing sophistication of cybercriminal activities creates significant challenges for traditional security systems, which rely heavily on unimodal detection methods. In this paper, we propose an integrated multi-modal deep learning methodology for better travel predictions by considering time series ordering od behavioral event transaction patterns with network traffic characteristics and textual threat intelligence. However, the suggested framework has modality-specific encoders which include Bidirectional Long Short-Term Memory; for behavioral sequence modeling, Graph Neural Networks, which focus on network topological processing, and BERT-based transformers with a focus on textual feature extraction. A new attention-based fusion mechanism finds latent representations from each modality, allowing the model to adjust source information dynamically depending on its predictive utility. Using three benchmarked datasets, IEEE-CIS Fraud Detection, CICIDS2017, and curated threat intelligence corpora for experimental evaluation of the proposed approach yield an F1-score of 0.912 and ROC-AUC of 0.947 and surpasses the respective single-modalities baselines by margins of 8.3% and 11.7%.