From Academic Research to Industrial Anti‑Fraud: Leveraging LLMs, Reinforcement Learning, and Model Distillation for Advertising Risk Detection

Collision of the Ivory Tower and Industry

During her studies at Tsinghua University, Xiaoting participated in data‑mining competitions, initially believing that high model metrics alone could solve problems; she achieved over 95% true‑positive rates with less than 0.35% false‑positive rates in anomaly detection.

After graduating, she joined JD.com’s retail technology team, confronting real‑world ad‑fraud during massive sales events where laboratory‑perfect metrics proved insufficient against massive, chaotic traffic.

Technical Detective: Using AI to Crack Black‑Market Codes

The CPS incentive model, while encouraging legitimate promotion, also attracted black‑gray market fraudsters who embed covert codes in address fields to claim commissions. Traditional regex filters failed to keep up with evolving patterns.

To address this, Xiaoting introduced large language models (LLMs) fine‑tuned with LoRA on a few thousand manually labeled anomalous addresses, enabling semantic understanding beyond surface text.

She further built a reinforcement‑learning‑with‑human‑feedback (RHLF) loop, where the model’s incorrect predictions are corrected by experts and fed back into training, improving generative recognition of hidden codes such as "78910" or "ATTTT233".

After three iterative versions, the system achieved a false‑positive rate of 0.3% and could accurately detect both explicit and covert address codes.

Not the Flashiest Tech, but the Most Effective Solution

Facing JD’s massive traffic, real‑time inference with a full‑scale LLM was impractical. Xiaoting employed knowledge distillation: the large model acted as a "senior professor" while a smaller model learned its expertise, achieving a balance of precision and latency after ten development cycles.

She emphasizes that true innovation lies in aligning cutting‑edge research with business constraints, selecting tools wisely, and understanding when to apply them.

Three Core Insights

1. Cost‑aware technology selection – evaluate data scale, compute cost, and business value; quantify how a 1% accuracy gain translates to annotation cost or how a 10 ms latency reduction impacts fraud interception.

2. Continuous knowledge evolution – stay abreast of advances like Attention mechanisms, filter relevant research, and adapt them to improve large‑model‑driven detection.

3. Cross‑domain thinking – use game‑theoretic models to anticipate attacker strategies, incorporate stochastic and nonlinear feedback mechanisms inspired by dissipative structures to create adaptive, self‑evolving defenses.

Conclusion

Technical professionals should cultivate "T‑shaped" abilities: deep vertical expertise combined with broad horizontal insight, enabling them to bridge academic breakthroughs and industrial impact while maintaining ethical standards.

Join Us

JD.com’s retail algorithm team is recruiting both campus and experienced talent for roles in CV, NLP, recommendation systems, reinforcement learning, large models, multimodal AI, AI infrastructure, operations research, and inference acceleration.

Apply via the JD recruitment portal or scan the QR codes provided.