How Large Models Are Revolutionizing Anti‑Fraud Detection: Fusion, Script Recognition, and Automated Iteration

Overview

The discussion focuses on applying large‑model AI to anti‑fraud (anti‑scam) scenarios, covering three core projects—multimodal fusion models, large‑model script recognition, and an automatic iteration framework—to address recall, adversarial robustness, and cross‑scene adaptability, followed by agent‑based efficiency improvements.

1. Evolution of Risk‑Control Technology

A real‑world telecom‑fraud case is dissected to illustrate the full attack chain: account registration, credential verification, embedding malicious links in platform posts, and luring users to external platforms. The case highlights three fraud characteristics: deep integration into complex business scenes, strong concealment, and rapid adversarial evolution.

2. Recognition Innovation: Large Models as Anti‑Fraud Weapons

2.1 Multimodal Fusion Model

The first project merges behavior vectors, text, and graph‑relation features into a single representation. Four fusion strategies were evaluated; attention‑based fusion (cross‑modal attention built on a Transformer encoder) was selected as the most effective. Experiments show that the attention‑fusion model improves accuracy by ~50 % and recall by ~63 % over simple feature concatenation, and outperforms traditional behavior and text models by ~31 % in both metrics.

2.2 Large‑Model Script Recognition

This project treats fraud scripts (the “playbook” of scammers) as classification tasks. Training data are JSON records containing instruction, input (the scam story), and output (the label). Prompt design emphasizes clear boundaries, parameterization for iterative optimization, and few‑shot examples for each script type (e.g., lead‑generation, rental‑assist, Hong‑Kong‑outsourced). The base model selection compared Qianwen‑34B and Qianwen‑37B; the 34B model was chosen for its balance of inference latency and accuracy, crucial for real‑time interception. Fine‑tuning follows a standard instruction‑tuning pipeline, and evaluation confirms superior accuracy and robustness compared with behavior‑only and text‑only models.

2.3 Automatic Iteration Framework

To keep pace with fast‑evolving fraud tactics, an end‑to‑end automated pipeline—sample acquisition → model training → evaluation → deployment—was built. Two parallel strategies run online: a “champion” model serving live traffic and a “challenger” model undergoing continuous A/B evaluation. When a challenger (e.g., version V8) surpasses the champion (V7) on predefined metrics, it automatically replaces the champion. The framework supports both traditional ML models (e.g., XGBoost with fixed architecture and iterative feature updates) and large models (prompt‑driven iteration and automated fine‑tuning). Experiments across behavior, text, and large models show that accuracy ratios remain stable over many iteration cycles, demonstrating mitigated model decay.

3. Efficiency Exploration: Agent‑Based Security Assistance

3.1 Data‑Query Agent

The agent receives natural‑language security queries (e.g., “刷单”, “杀猪盘”) and generates corresponding SQL statements to retrieve data from internal tables. A RAG (Retrieval‑Augmented Generation) knowledge base supplies table metadata, fraud definitions, and alias mappings. Prompt engineering is iteratively refined to enforce strict output formatting, ensuring reliable SQL generation. Human evaluation shows a monotonic increase in correctness as prompts evolve.

3.2 Strategy‑Generation Agent

Building on the data‑query capability, the strategy‑generation agent constructs end‑to‑end fraud‑mitigation policies. The workflow includes: (1) data retrieval, (2) user‑profile linking, (3) analytical chain construction, and (4) policy output. The agent mimics expert reasoning by interviewing frontline operators, extracting their analysis steps, and reproducing them. Implemented on the low‑code Dify platform, the system comprises ~200 workflow nodes and produces JSON‑formatted policies that can be directly consumed by the live risk‑control system.

3.3 Agent Evaluation and Limitations

Human scoring of generated SQL and policies confirms continuous improvement with prompt iterations. However, limitations remain: the agents excel mainly in text‑centric fraud scenarios, struggle with behavior‑heavy cases, and their performance scales with model size and GPU resources, affecting real‑time latency.

Conclusion

On the recognition side, multimodal fusion, script‑recognition fine‑tuning, and automated iteration collectively resolve recall gaps, adversarial resistance, and cross‑scene adaptability, delivering higher precision and stability. On the efficiency side, data‑query and strategy‑generation agents achieve near‑human performance in selected tasks, dramatically reducing manual effort and accelerating fraud‑mitigation workflows. Remaining challenges include expanding agent reasoning breadth and optimizing resource consumption for large‑model deployment.