Unsupervised Anomaly Detection for SIM Card Fraud Using Autoencoder-Based Reconstruction Error Analysis

Detecting commercial fraud in telecommunications networks remains critically challenging as sophisticated schemes systematically evade traditional rule-based detection systems. This paper introduces an unsupervised deep learning framework for SIM card fraud detection, rigorously evaluated on a real-world dataset comprising 1.5 million commercial transactions. From eight raw transactional attributes, we engineer a discriminative feature space of 26 behavioral and temporal indicators. We systematically evaluate multiple autoencoder architectures "shallow, medium, deep, and denoising variants" and a Variational Autoencoder (VAE), benchmarking them against a One-Class Support Vector Machine (OCSVM) baseline. Evaluation on a labeled test set of 291000 transactions reveals that the Deep Autoencoder achieves superior performance with an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.885, substantially outperforming all competing approaches. Beyond raw detection accuracy, our analysis establishes that the model excels at riskbased transaction ranking, providing an effective prioritization mechanism for human investigators. We integrate t-distributed Stochastic Neighbor Embedding (t-SNE) for latent space visualization, SHapley Additive exPlanations (SHAP) for model interpretability, and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) for anomaly characterization, uncovering over 4000 distinct anomaly typologies. This framework delivers not merely alerts, but actionable, interpretable intelligence into the operational patterns underlying fraudulent activities.