FRAUDSkill: Structured Frozen-Weight Skill Optimization for Audio Anti-Fraud Detection

FRAUDSkill. Structured Frozen-Weight Skill Optimization for Audio Anti-Fraud Detection. By Chengxian Hu and eleven others, from the People's Public Security University of China, J D Technology, and four other institutions. Published on the sixteenth of September. No code repository is linked to it.

Large audio-language models show promise for anti-fraud detection, processing speech directly and reasoning over evidence of fraud. Deploying them brings requirements of its own. Predictions have to stay inside a predefined label space and follow a structured decision protocol: identify the service scenario, detect fraud, and, only when fraud is found, classify its type. Existing fine-tuning and prompt-based approaches tend to encode task knowledge, constraints, and decision rules into model parameters or into prompts maintained by hand. Either way, they are hard to adapt as fraud patterns and labeling policies change.

FRAUDSkill keeps the audio-language model frozen and unchanged. What gets optimized instead is an external layer: skill programs, policies specific to each route, and decision rules. Task knowledge moves out of the weights and into something that can be revised. When fraud patterns or labeling policies change, it is that external layer that gets updated, not the model beneath it. The authors call this structured frozen-weight adaptation: adaptation happens, but the weights stay fixed.

Two further pieces keep outputs inside the protocol. Structured output control governs the form of what the model returns, and it is combined with multi-path inference guided by validation. Together they are there to ensure that predictions comply with the protocol, meaning that each prediction falls in the predefined label space and follows the three-step order of scenario, fraud, and fraud type. An answer outside the label space counts as an invalid output, which is why the rate of invalid outputs is reported next to accuracy.

On the TeleAntiFraud benchmark, FRAUDSkill reaches seventy-three point five percent macro-F1. That outperforms the shared frozen-model baseline by thirty- one point nine six percent, while invalid outputs fall to one point nine four percent. Extensive experiments support external skill optimization as an effective and adaptable approach to structured audio anti-fraud detection.

The model is left exactly as it was. Everything that encodes the task, and everything likely to change as fraud evolves, lives outside it, where it can be optimized and replaced without touching the weights.