https://dl.acm.org/doi/epdf/10.1145/3637528.3671534 cyber fraud

Unique Title: “CyShield: Real-Time Explainable Threat Detection in Financial Transaction Networks Using Masked Heterogeneous Graph Learning” Abstract: Financial cyberfraud remains a persistent threat to digital banking ecosystems, necessitating detection systems that are not only accurate but also interpretable for security analysts. Current graph-based fraud detection models often operate as black boxes, limiting their adoption in regulated environments where explainability is mandatory. This paper introduces CyShield, a novel self-interpretable detection framework that integrates adaptive masking mechanisms within a heterogeneous graph transformer to simultaneously identify and explain fraudulent activities. The system learns dual-attention masks—over node attributes and relational edges—during model training, enabling real-time explanation generation without post-hoc overhead. Evaluated on real-world financial data from a major banking institution, CyShield achieves a 99.69% AUC and 99.38% recall, outperforming state-of-the-art detectors while providing interpretable evidence for flagged transactions. Additionally, it reduces explanation inference time to 0.4 milliseconds per case—over 1000× faster than existing explainers—making it viable for high-throughput, compliance-sensitive cybersecurity operations