DeepScrub: LLM-Based Traceable Fraud Detection for O2O Platforms
A new framework named DeepScrub has been developed by researchers, utilizing reinforcement learning and large language models (LLMs) to detect fake-order fraud on extensive online-to-offline (O2O) service platforms. This approach overcomes the shortcomings of current techniques that depend on manually crafted features and yield opaque decisions. DeepScrub features three significant advancements: a semantic unification module that translates diverse risk signals into LLM-friendly textual formats; ongoing pre-training on risk-control datasets to embed domain expertise, with task rewards assessing both prediction accuracy and reasoning quality; and the SUggest-REflect (SURE) mechanism, which combines expert insights with model self-evaluation for continuous improvement. The system was validated on a genuine fake-order fraud detection dataset, showcasing clear reasoning in its fraud detection outcomes.
Key facts
- DeepScrub is a reinforcement learning framework built upon LLMs for fake-order fraud detection.
- It addresses challenges of large O2O service platforms.
- Existing approaches rely on expert-designed features and produce black-box decisions.
- Semantic unification module converts heterogeneous risk signals into textual descriptions.
- Continued pre-training on risk-control corpora injects domain knowledge.
- Task rewards jointly evaluate prediction correctness and reasoning quality.
- SURE mechanism incorporates expert feedback and model self-checking.
- Tested on a real-world fake-order fraud detection dataset.
Entities
Institutions
- arXiv