Real-Time Fraud Detection in Telecom Charging Systems Using AI

Abstract

Building upon the proposed VAE‑GAN architecture, our system incorporates dynamic feature enrichment by integrating device fingerprinting and behavioral analytics into the anomaly detection pipeline. Specifically, device-specific metadata such as SIM usage patterns, handset IMEI clustering, and network access behaviors are continuously fed into the VAE encoder, enabling the model to learn a richer representation of “normal” subscriber activity. The adversarial component (GAN discriminator) then sharpens the detection boundary by contrasting these learned patterns against synthetically generated anomalies. Such a dual-encoder framework not only improves the detection of subtle fraud variants like SIM box bypass or IRSF, but also enables the framework to adapt in-stream as fraudsters shift tactics. Preliminary results on CDR datasets show a marked increase in precision (up to ~93%) and recall (~90%) outperforming conventional autoencoder‑only models while maintaining low latency suitable for real-time telecom charging environments . Moreover, by leveraging continuous retraining on incoming call streams, the system exhibits resilience to concept drift, ensuring sustained performance in dynamic telecom ecosystems