Applying Artificial Intelligence to Cross‑Border Risk Control: Practices and Insights

We are XiaoDun Technology, a sub‑brand of TongDun, focusing on cross‑border risk control solutions.

Background: With domestic traffic saturation, many merchants are expanding overseas, encountering new risks that differ from domestic ones. Our servers are located in North America, Europe, Singapore, and Indonesia.

Customers: We provide SaaS services to a wide range of merchants and platforms, aiming to integrate diverse data sources for risk analysis.

Intelligent Risk‑Control System – Case Study: Traditional rule‑based address filtering struggles with countless address variations. We convert address variants into feature vectors and compute similarity in a vector space, enabling more effective detection compared with static rules.

Differences from Traditional Risk Control: Rule‑based systems excel at quick, known‑scenario mitigation but struggle with complex, unstructured data such as text and images. Our algorithmic approach leverages richer data and advanced models to handle these cases.

Layered Architecture: The system consists of four layers – data collection, data processing (standardization and storage), algorithm development (model training on processed features), and application layer (real‑time or offline decision engine integrated with existing rule engine).

Risk‑Control Practice – Industry Background: Overseas merchants (e‑commerce, live‑streaming, entertainment) face new fraud types such as account abuse, payment refusals, and marketing abuse.

Transaction Risk: Includes charge‑backs (refusals) due to credit‑card payments and in‑app purchases (e.g., virtual goods). We analyze user behavior sequences (clicks, browsing, add‑to‑cart) to differentiate legitimate from malicious patterns.

Our modeling pipeline transforms event sequences into embeddings, then applies algorithms such as CNN, LSTM, and Transformer. This generic pipeline allows us to serve many customers with minimal custom feature engineering.

Marketing Risk: Similar to domestic scenarios (e.g., coupon abuse, fake flash sales) but amplified overseas. We build a large graph of user, device, IP, and activity nodes, applying GNNs to compute risk scores for each node and community.

Our graph‑based approach can identify about 30% more fraudulent cases compared with traditional supervised learning or rule‑based methods.

Cross‑Border Risk Reflections: Value lies in handling soft, non‑rule‑based risks; challenges include sparse or inconsistent data samples across merchants and the need for explainability in a negative‑impact domain.

We aim to provide a flexible, semi‑customizable platform that balances generic capabilities with merchant‑specific tuning, reducing overall risk by an estimated 10‑15%.

Q&A: Discussed differences between domestic and overseas payment fraud, address vectorization for clustering and real‑time interception, and brief notes on overseas money‑laundering techniques.

Thank you for attending.