Skip to content
Book Open access

Every Victim Tells a Dialogue: Reconstructing 80K Audio-Text Anti-Fraud Conversations from Case Narratives

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 43 references

Abstract

As telecom fraud increasingly threatens global financial security, the demand for reliable detection models has surged. Advancing these models, however, is severely bottlenecked by two critical challenges: (i) the limited availability of real-world data due to strict privacy regulations on scam recordings; and (ii) poor annotation quality caused by unstructured formats and unstandardized categories. While existing works leverage LLMs and hand-crafted prompts to synthesize fraudulent dialogues, the resulting datasets often lack grounding in documented fraud cases, increasing the risk of factual and procedural drift. To address these challenges, we propose Narrative-to-Dialogue Reverse Reconstruction (NDRR), a framework for constructing case-grounded synthetic fraud dialogues from real-world case narratives. Through three stages—Fraud Procedure Extraction, Fact-Anchored Dialogue Reconstruction, and Consistency Verification—we construct FraudRecon, a large-scale dataset of 80,319 paired audio-text dialogues covering scam and matched benign interactions. It features comprehensive three-level annotations across 12 fraud categories, 6 scam-specific procedural phases, and 42 interaction signals. We also introduce FraudRecon-Bench, a five-task benchmark encompassing Fraud Detection, Category Classification, Phase Identification, Signal Recognition, and Evidence-Chain Reasoning. Extensive experiments on 13 LLMs (11 zero-shot and 2 fine-tuned) demonstrate that FraudRecon-Bench poses a significant challenge: the best zero-shot overall score is 53.3/100, while fine-tuning on FraudRecon improves the overall score by up to 32.8 points on this benchmark. Our code and data-access instructions are available at https://github.com/Henryhe09/FraudRecon.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.