Synthetic Identity Fraud (SIF) is an emerging financial threat where fraudsters construct fictitious identities by blending legitimate and fabricated Personally Identifiable Information (PII). Detecting SIF is particularly challenging due to its low frequency, evolving structure, and similarity to genuine customer profiles, rendering conventional fraud detection systems ineffective. To address this gap, we propose a hybrid machine learning framework combining symbolic PII transformation, behavioral profiling, anomaly scoring, and ensemble classification in a layered pipeline. Our method integrates unsupervised anomaly detection with supervised learning, incorporating explainability via SHAP values and robustness through Monte Carlo Dropout. Experiments on a real-world financial dataset demonstrate strong performance, achieving an F1-score of 0.822, precision of 0.85, recall of 0.79, and AUC-PR of 0.88. The framework is designed to operate effectively in highly imbalanced and operationally constrained environments, offering a practical solution for early SIF detection in financial systems.
Khan et al. (Mon,) studied this question.