Are You for Real? Detecting Identity Fraud via Dialogue Interactions

Weikang Wang, Jiajun Zhang, Qian Li, Chengqing Zong, Zhifei Li

Abstract

Identity fraud in financial services is a significant challenge, as fraudsters often perfectly memorize fake personas. This paper introduces an interactive dialogue system designed to detect such fraud by probing "contextual knowledge." By asking "derived questions"—facts related to an applicant's personal information (e.g., nearby landmarks)—the system distinguishes between legitimate users and fraudsters. Utilizing a Knowledge Graph (KG) and hierarchical reinforcement learning, the proposed model significantly outperforms standard rule-based methods in both accuracy and efficiency.

Introduction

Identity fraud in loan applications represents a massive financial drain. Traditional detection methods often rely on static verification or direct questioning. However, modern fraudsters are increasingly sophisticated, often memorizing a "fake persona" (name, address, company) perfectly. While a fraudster might know a fake address, they likely lack the intuitive contextual knowledge of the neighborhood or the claimed workplace. This paper addresses the lack of automated methods to dynamically probe this "contextual knowledge" through natural conversation.

System Architecture and Methodology

The authors propose a two-part system to generate and manage these interrogative dialogues.

1. Knowledge Graph (KG) Constructor

To generate intelligent questions, the system links personal information entities (like a specific university or company) to existing Chinese KGs and geographic data from Amap. This allows the creation of triplets such as (Nanjing University, FoundedDate, 1902) or (Company X, NearestSubway, Station Y).

2. Dialogue Management

The "brain" of the system consists of two sub-modules:

  • KG-based Dialogue State Tracker (KG-DST): Uses message passing on the graph to update the conversation state. Correct answers about local landmarks increase the system's "belief" in the applicant's legitimacy.
  • Hierarchical Dialogue Policy (HDP): Employs a two-level approach where a Manager agent decides which category to verify (e.g., "School"), and a Worker agent selects the specific derived question from the KG.
The hierarchical dialogue management system architecture, showing the interaction between the Manager, Worker, and Knowledge Graph.

The system is trained using Reinforcement Learning (RL) to maximize detection accuracy while minimizing the number of questions asked.

Key Insights

  • The "Derived Question" Strategy: Fraudsters typically fail on facts related to their fake persona. Real residents know nearby landmarks intuitively; fraudsters do not.
  • Hierarchical RL: Splitting the task into a Manager (strategy) and Worker (execution) makes RL training stable and efficient despite a massive action space.
  • Internet "Spread Degree": The system uses search engine hit counts to estimate how "knowable" a fact is. Obscure facts known by a user are strong positive signals, while unknown common facts are red flags.

Experimental Results

The system was evaluated against several baselines, including rule-based systems. The "Full-S" model, which incorporates the full hierarchical RL and KG-DST framework, demonstrated superior performance.

Comparison of Accuracy and Average Turns across different models. The Full-S model achieves the highest accuracy (0.884).

As shown in the results, the proposed model achieved 88.4% accuracy, significantly outperforming the "Flat Rule" (74.8%) and "Hierarchical Rule" (78.7%) baselines. Furthermore, the system learned to prioritize high-impact categories like "School" and "Company," where fraudsters are statistically more likely to lie to secure larger loans.

Limitations and Future Work

While effective, the system faces challenges such as real-time "googling" by applicants, which the authors suggest mitigating via response time limits. Additionally, Knowledge Graph sparsity remains an issue for less-documented entities. Future work involves improving KG density and exploring deployment in high-fidelity simulators to navigate privacy regulations.

Reference: Wang, W., Zhang, J., Li, Q., Zong, C., & Li, Z. (2019). Are You for Real? Detecting Identity Fraud via Dialogue Interactions. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP).

Summary updated

BibTeX
@inproceedings{wang2019are,
  author = {Weikang Wang and Jiajun Zhang and Qian Li and Chengqing Zong and Zhifei Li},
  title = {{Are You for Real? Detecting Identity Fraud via Dialogue Interactions}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  doi = {10.18653/v1/d19-1185},
}
Figure