Unsupervised TTL-Based Deep Learning for Anomaly Detection in SIM-Tagged Network Traffic

Abstract

The rise of SIM cloning, identity spoofing, and covert manipulation in mobile and IoT networks has created an urgent need for continuous post-registration verification. This work introduces an unsupervised deep learning framework for detecting behavioral anomalies in SIM-tagged network flows by modeling the intrinsic structure of benign behavioral descriptors (TTL, timing drift, payload statistics). A Temporal Deep Autoencoder (TDAE) combining Conv1D layers and an LSTM encoder is trained exclusively on normal traffic and used to identify deviations through reconstruction error, enabling one-class (label-free) training. For deployment, alarms are set using an unsupervised quantile threshold τα calibrated on benign traffic with a false-alarm budget; τ* is reported only as a diagnostic reference for model comparison. To ensure realism, a large-scale corpus of 3.6 million SIM-tagged flows was constructed by enriching public IoT traffic with pseudo-operator identifiers (synthetic SIM tags derived from device identifiers) and controlled anomaly injections. Cross-domain experiment transfer under SIM-grouped protocol: Training on clean Cassavia-like traffic and testing on attack-rich Guarascio-like flows yields a PR-AUC of 0.93 for the proposed Conv-LSTM Temporal Deep Autoencoder, outperforming Dense Autoencoder, Isolation Forest, One-Class SVM, and LOF baselines. Conversely, the reverse direction collapses to PR-AUC ≈0.5, confirming the absence of data leakage and the validity of one-class behavioral learning. Sensitivity analysis shows that performance is stable around the unsupervised quantile operating point. Overall, the proposed framework provides a lightweight, interpretable, and data-efficient behavioral verification layer for detecting cloned or unauthorized SIM activity, complementing existing registration mechanisms in next-generation telecom and IoT ecosystems.