Design Principles for AI‑Driven Anti‑Fraud Systems

This article, based on Tan Chilie’s presentation at the 2018 AI Pioneer Conference, introduces Tongdun Technology, a provider of intelligent risk control and anti‑fraud services for finance, O2O, and other industries, highlighting its expertise in blockchain, big data, machine learning, and AI.

It describes the high‑risk problems faced by anti‑fraud systems, such as coupon‑grabbing ("薅羊毛"), account theft, fake registrations, card fraud, and coordinated attacks by large fraud gangs that exploit vulnerabilities across multiple channels.

The proposed anti‑fraud solution combines real‑time and offline monitoring, intelligence gathering, and both supervised (classification, regression) and unsupervised methods (graph mining, clustering) to detect and intercept fraudulent activities, emphasizing the need for multi‑dimensional feature analysis.

The AI design principles are presented in three parts: (1) Early warning and risk discovery through real‑time and offline monitoring, reliable intelligence, time‑series and trend analysis, and statistical anomaly detection; (2) Enriching models with diverse, multi‑dimensional features such as event correlation, geographic data, user behavior, and persona profiling; (3) Human‑machine collaboration, where rule‑based decisions, expert analysis, and supervised learning work together with unsupervised techniques to create a layered, adaptive defense system.

The article concludes that confronting fraud requires continuous iteration, cross‑industry collaboration, and a multi‑level intelligent risk control architecture that integrates business rules, supervised models, and unsupervised learning to stay ahead of evolving threats.

Author: Tan Chilie, Director of Anti‑Fraud Algorithms at Tongdun Technology.