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

Introduction – The talk introduces the application of artificial intelligence in cross‑border risk control, focusing on practical deployment. It outlines three main parts: who the team is, the intelligent risk‑prevention system, and cross‑border risk‑control practice.

Who We Are – The speaker represents a sub‑brand of Tongdun Technology, dedicated to overseas risk management. Their infrastructure spans North America, Europe, Singapore, and Indonesia, serving a diverse set of SaaS customers.

Background & Customers – With domestic traffic plateauing, many merchants are expanding overseas, encountering new risk types. The company collects data from various regions and serves many merchants, aiming to provide a unified SaaS risk‑control platform.

Intelligent Risk‑Control System

1. Case Example – Standard addresses can mutate into countless variants, making rule‑based detection difficult. By converting address variants into feature vectors and measuring similarity in a latent space, the system can effectively block related fraud.

2. Differences from Traditional Rule‑Based Systems – Traditional systems rely on static lists and rules, suitable for simple scenarios but ineffective for complex or unstructured data such as text and images. The AI‑driven approach leverages richer data and algorithmic models, though it depends heavily on client data quality.

3. Layered Architecture – The architecture consists of four layers: (a) data collection from multiple platforms, (b) data processing and feature standardization, (c) algorithm development (CNN, LSTM, Transformer, GNN), and (d) application layer where decisions are made via a decision engine that combines AI models with existing rule sets.

Intelligent Risk‑Control Practice

1. Industry Background – Overseas merchants (e‑commerce, live‑streaming, entertainment) face new risks: identity fraud, chargebacks, and marketing abuse. The business security is divided into identity, transaction, and marketing risks.

2. Transaction Risks – Two major categories: chargebacks (common with credit‑card payments abroad) and in‑app purchases (especially in gaming). Both are modeled using user behavior sequences, embedding events, and contextual features, then processed with CNN/LSTM/Transformer models to detect anomalies.

3. Data‑Sequence Modeling – Event sequences (clicks, views, add‑to‑cart) are enriched with contextual features and passed through embedding layers, followed by deep models. This approach yields 2‑3× higher recall compared to traditional machine‑learning pipelines while keeping integration costs low.

4. Marketing Risks – Similar to domestic marketing abuse, overseas campaigns suffer from “coupon‑clipping” and fraudulent referrals. The team builds a large‑scale graph of users, devices, and IPs, applying Graph Neural Networks (GNN) to detect suspicious clusters, achieving ~30% higher fraud detection than rule‑based methods.

5. Cross‑Border Risk Reflections – Challenges include varying compliance regulations, limited and heterogeneous data samples, and the need for model interpretability. The solution balances generic models with customizable components to serve diverse merchants while maintaining explainability.

Q&A Highlights

• Differences between domestic and cross‑border fraud stem from regulatory diversity and the prevalence of credit‑card payments.

• Address vectors are used both for offline clustering to discover fraud rings and for real‑time similarity checks in the rule engine.

• Overseas money‑laundering tactics are similar to domestic ones but leverage region‑specific channels; data‑driven monitoring is essential.

The session concludes with a summary of the AI‑driven risk‑control framework and its impact on reducing fraud in international commerce.