How AI and LLMs Power JD’s Real-Time Advertising Anti‑Fraud System

Collision of Academic and Industry Thinking

During her studies at Tsinghua University, the author participated in data‑mining competitions, focusing on optimizing models for anomaly detection and achieving high precision and low false‑positive rates, which fostered a strong belief in algorithmic solutions.

After graduating, she joined JD.com’s retail technology team, confronting real‑world advertising fraud during massive traffic spikes. Laboratory‑perfect metrics proved insufficient, prompting a shift toward continuously evolving solutions that balance technical feasibility, business value, and implementation cost.

Tech Detective: Using AI to Crack Black‑Market Codes

The CPS incentive model in advertising attracts black‑gray market fraud, where malicious actors embed hidden codes in address fields to claim commissions. Traditional regex‑based filters fail to adapt to evolving patterns.

Introducing large language models (LLMs) into the detection pipeline enables semantic understanding of address texts. By fine‑tuning open‑source LLMs with LoRA on a few thousand manually labeled anomalous addresses, the model learns to recognize hidden patterns.

A reinforcement learning framework (RHLF) incorporates human feedback to correct model errors, allowing the system to iteratively improve its generative recognition of coded addresses.

Examples such as "3栋78910单元1023室" and "3栋2单元1023室ATTTT233" demonstrate the model’s ability to identify concealed numeric strings that traditional regex cannot detect.

After three rounds of iteration, the system reduced the false‑positive rate to 0.3% while accurately capturing both explicit and covert codes, showcasing the power of combining LLMs with domain‑specific fraud detection.

To meet JD’s massive traffic demands, model distillation was employed: a large teacher model transfers its knowledge to a lightweight student model, achieving a balance between precision and latency suitable for real‑time inference.

Final Thoughts

Effective engineers need a “T‑shaped” skill set—deep vertical expertise combined with broad horizontal insight—to evaluate models not only by accuracy but also by data scale, computational cost, and business impact.

Continuous learning of cutting‑edge research (e.g., attention mechanisms) and applying them to practical systems bridges the gap between academia and industry.

Cross‑domain thinking, such as using game theory to anticipate fraud patterns and incorporating stochastic feedback mechanisms, equips teams to stay ahead of evolving black‑market tactics.