The Synthetic Identity Landscape: A Canonical Survey of Deep Learning Methodologies and Research Frontiers

Abstract

This research provides a comprehensive and canonical systematic review of the evolution of deep learning methodologies for detecting synthetic identity fraud. As synthetic generation techniques have progressed from simple rule-based heuristics to sophisticated Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), detection strategies have undergone a parallel transformation. This paper categorizes the existing literature into four primary technical taxonomies: supervised feature-based detection, unsupervised anomaly discovery, graph-based relational analysis, and multimodal latent fusion. The survey evaluates the performance, scalability, and robustness of state-of-the-art models against known adversarial attack vectors across multiple sectors, including finance and digital biometrics. By synthesizing findings from over 200 high-impact studies, the research identifies critical research frontiers, specifically regarding model explainability, adversarial robustness, and the impact of data sparsity on model training. This work serves as a foundational reference for academic researchers and industry practitioners, providing a structured roadmap for the next generation of identity protection research and establishing a baseline for future scholarly inquiry.