AI-based Identity Fraud Detection: A Systematic Review

Chuo Jun Zhang, Asif Q. Gill, Bo Liu, Memoona J. AnwararXiv 2025 · DOI · arXiv

Introduction

The digital transformation has turned Personally Identifiable Information (PII) into a high-value target for cybercriminals. Traditional security measures are increasingly vulnerable to Generative AI (GAI), which enables the creation of sophisticated deepfakes and synthetic identities at scale. This systematic literature review (SLR), conducted by researchers at the University of Technology Sydney, maps the evolving AI-driven arms race in identity fraud by analyzing 43 core studies published between 2020 and 2024.

The Identity Fraud Landscape

Identity fraud extends beyond simple financial theft; it involves the fabrication or theft of entire personas. The "Identity Fraud Chain" illustrates how data stolen from various sectors—including healthcare, government, and financial institutions—is processed by malicious AI to create synthetic or fabricated identities used for a wide range of crimes, from financial fraud to illegal immigration.

The Identity Fraud Chain: From data intrusion to fraudulent activities.

Methodology and Taxonomy

The authors filtered an initial pool of 1,257 papers down to 43 high-quality studies to develop a novel taxonomy for identity fraud detection. This framework categorizes defenses into two distinct operational phases:

  1. Authentication (The Gatekeeper): Focuses on initial verification using biometrics (face, fingerprint, voice) and visual anomaly detection to identify deepfakes or tampered documents.
  2. Continuous Authentication (The Sentry): Employs User and Entity Behaviour Analytics (UEBA) to monitor sessions post-login, flagging deviations in historical patterns like keystroke dynamics, touchscreen gestures, or transaction locations.
Taxonomy of AI-based identity fraud detection methods.

Key Insights and Results

  • Deep Learning Dominance: Deep Learning (DL) models, such as CNNs (XceptionNet, ResNet), power 76% of visual anomaly detection and 53% of behavioral analysis, often achieving over 95% accuracy in controlled environments.
  • Shift to UEBA: Continuous monitoring is becoming essential. Social features and communication styles (writing style, tone) are now critical components of resilient user profiling.
  • Biometric Vulnerability: While facial recognition is the most researched biometric (appearing in 30% of studies), it is also the primary target for GAI-driven "morphing" and presentation attacks.
  • Research Distribution: Currently, 55.8% of research focuses on initial authentication, while 44.2% addresses the growing need for continuous authentication.

Challenges and Future Directions

The field faces significant hurdles, including the "Privacy Paradox"—the need for high-quality data to train models versus strict privacy regulations like GDPR. Additionally, many high-performing DL models are computationally heavy, hindering real-time deployment on mobile devices.

Future research is trending toward unsupervised learning for "zero-day" fraud detection (identifying previously unseen tactics) and multimodal integration, which combines face, voice, and behavioral data into a single, unforgeable identity score.

Summary updated

BibTeX
@misc{zhang2025ai,
  author = {Chuo Jun Zhang and Asif Q. Gill and Bo Liu and Memoona J. Anwar},
  title = {{AI-based Identity Fraud Detection: A Systematic Review}},
  year = {2025},
  eprint = {2501.09239},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.09239},
}
Figure