Merchant Security: Authenticity Recognition and Transaction Risk Identification Using AI Techniques

The talk introduces a three‑stage merchant security framework addressing authenticity recognition and transaction risk identification, followed by intelligent decision making.

1. Merchant Authenticity Recognition

Authenticity is evaluated from three dimensions: credential authenticity, text authenticity, and transaction authenticity.

Credential authenticity: Images such as storefront photos and business licenses are verified using self‑supervised learning (SimCLR) to reduce labeling costs, and semi‑supervised methods (FixMatch) with domain classification loss to handle diverse credential types and complex backgrounds.

Text authenticity: Short textual fields (merchant name, address, industry) are processed with a Bi‑LSTM based short‑text pre‑training model for classification, and SimCSE for text matching, enabling detection of batch‑signing risks.

Transaction authenticity: Transaction patterns, timestamps, and LBS data are analyzed to confirm real business activity, infer operating scenarios, and validate declared industry information using explainable boosting models and Isolation Forest for anomaly detection.

These three analyses answer the fundamental questions of who the merchant is, where they operate, and what industry they belong to.

2. Merchant Transaction Risk Identification

Risk detection during the transaction phase leverages heterogeneous sequence algorithms and graph sparsification techniques.

Heterogeneous sequence algorithms: Buyer and merchant event sequences are embedded and fed into Transformers; local‑attention fusion of buyer and merchant embeddings improves AUC by ~3%.

Graph sparsification algorithms: To handle large‑scale heterogeneous graphs, relationship‑specific sparsification is performed using attention‑based node scoring, followed by aggregation via a HAN‑style multi‑relation graph to obtain robust node embeddings, reducing computation while improving coverage by 6‑10%.

3. Merchant Intelligent Decision

The final decision layer balances risk mitigation and user experience, addressing challenges such as model iteration speed, data bias, and explainability. Incremental updates via reinforcement‑learning‑style feedback, A/B testing for unbiased data, and causal inference for explainable controls are discussed.

Overall, the framework integrates AI methods across authenticity verification, risk detection, and decision making to protect digital commerce ecosystems.