Classifying Scam Calls Through Content Analysis With Dynamic Sparsity Top-<i>k</i> Attention Regularization

The rise of scam calls in recent years necessitates effective countermeasures against these fraudulent activities, which cause financial losses and threaten personal security. While previous research utilizing traditional machine learning techniques has fallen short in today’s technological landscape, this study introduces a novel approach for recognizing scam calls by analyzing call contents. Leveraging natural language processing techniques and deep learning methodologies, we propose the D-STAR (Dynamic Sparse Attention with Top-<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> Regularization) model, a transformer-based architecture designed to enhance scam call content detection. Unlike conventional models, D-STAR integrates Dynamic Sparse Attention (DSA), Top-<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> selection, and sparsity regularization, optimizing computational efficiency while preserving key scam-related contextual information. Our data set consists of 400 scam and 400 non-scam call transcripts, collected from publicly available sources such as social media, news reports, and discussion forums. To ensure dataset diversity, ChatGPT was utilized only to augment real scam scenarios across different contexts while preserving their core fraudulent structures. The model was evaluated using various hyperparameter configurations and managed to achieve an accuracy of 94%, a recall of 91.67%, and an F1-score of 84.98% in classifying scam call contents, outperforming state-of-the-art baselines such as CNN, LSTM, Decision Tree, Random Forest, and SVM in the scam call detection domain. A knowledge graph-based preprocessing technique was also introduced to enrich scam-related contextual understanding. The proposed approach demonstrates its effectiveness in enhancing scam call classification while maintaining computational efficiency. Future work will focus on real-world validation with telecom providers and further optimizations for real-time deployment.