Leveraging Bilateral Temporal Self-Attention (BilTSM) Networks for Real-Time Scam Call and Cyber Fraud Detection: Methods, Applications, and Performance Evaluation

Abstract

Scam phone calls and cyber fraud cause staggering financial losses worldwide, demanding advanced real-time detection solutions. This paper proposes a novel deep learning architecture, Bilateral Temporal Self-Attention (BilTSM) Network, which integrates bilateral (bi-directional) temporal context modeling with self-attention mechanisms and temporal shift operations for efficient sequential data analysis. We apply BilTSM to detect scam calls and fraudulent activities in cybersecurity (e.g. credit card transactions, network intrusions) and benchmark its performance against Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Transformer-based models, and Random Forest classifiers. The BilTSM model is evaluated on publicly available datasets, including the Kaggle Credit Card Fraud dataset, CTU-13 botnet traffic, and UNSW-NB15 network intrusion data. Our results show that BilTSM achieves superior accuracy, precision, recall, F1-score, and ROC-AUC compared to baseline models, while operating with low latency suitable for real-time deployment. We present a comprehensive theoretical background on the BilTSM architecture, including mathematical formulations of its bilateral self-attention and temporal shift modules. Detailed experimental settings, hyperparameter tuning, and data preprocessing strategies are described. We also demonstrate BilTSM’s practicality in real-world telecom infrastructure and cybersecurity operations, discussing deployment considerations. The paper concludes with insights into BilTSM’s advantages, current limitations (such as class imbalance and novelty detection challenges), and future research directions, paving the way for more robust real-time fraud detection systems.