Adversarial Identity Synthesis and Detection: A Systematic Survey of Foundation Model Architectures and Defensive Frontiers

Abstract

This research provides a systematic survey of the evolving landscape of synthetic identities in the context of large-scale foundation models. Traditional identity verification assumes synthetic personas exhibit detectable statistical anomalies; however, the emergence of generative foundation models has significantly lowered the barrier for creating high-fidelity, multimodal synthetic identities that evade current detection pipelines. The study categorizes recent advances in generative identity synthesis, including automated persona generation and deepfake-based biometric spoofing. Detection methodologies ranging from supervised neural classifiers to graph-based relational analysis are evaluated against these advanced generative threats. By synthesizing findings across several technical domains, the paper identifies critical vulnerabilities in existing verification architectures and proposes a roadmap for self-evolving, adaptive defensive frameworks. The analysis demonstrates that behavior-based provenance tracking, real-time adversarial monitoring, and multimodal consistency checks are essential for maintaining digital integrity. These perspectives provide a structured foundation for developing next-generation identity verification systems capable of resisting sophisticated, machine-generated synthetic identities.