[论文] Incremental Risk Assessment of Progressive Elder Financial Scams via I...

论文概要

研究领域: ML
作者: Parviz Ghafariasl, Weimin Fu, Xiaolong Guo
发布时间: 2026-09-03
arXiv: 2509.00004

中文摘要

针对老年人的金融诈骗越来越多地通过文本和语音渠道（如电子邮件、短信和电话）发生，这些诈骗在多个对话轮次中逐步展开，从冒充或随意接触开始，通过建立信任和制造紧迫感升级，最终以索取敏感信息或财务转账告终。由于风险信号在轮次间逐步显现，有效检测需要能够在资源受限的部署设置下持续更新风险估计的模型。我们提出了一个累积轮次风险评估框架，逐步聚合对话轮次并在每一步重新估计风险，实现对逐步演变的对话的动态诈骗监控。构建了一个多轮对话数据集，涵盖投资、慈善和技术支持诈骗场景，每个对话包含2到8个轮次，并在每个累积阶段标注定性风险等级、连续风险评分、解释性理由和安全建议。四个小型语言模型（Phi-4、LLaMA-3.2、DeepSeek-R1和Qwen3）在统一训练框架下进行微调并评估。微调后的小型模型在保持适合移动设备和资源受限部署设置的紧凑架构的同时，捕获了与欺诈相关的语言线索和跨轮次升级模式。在评估的模型中，Phi-4和LLaMA-3.2相对于其参数规模实现了更强的轮次感知风险估计性能。这些结果表明，结构化累积建模可以支持面向部署的增量诈骗风险评估，同时突出了紧凑语言模型在隐私感知和设备端欺诈保护方面的潜力。

原文摘要

Financial scams targeting older adults increasingly occur through text and voice channels such as email, SMS, and phone calls, unfolding over multiple conversational turns that begin with impersonation or casual contact, escalate through trust building and urgency, and culminate in requests for sensitive information or financial transfers. Because risk signals emerge incrementally across turns, effective detection requires models that continuously update risk estimates under resource-constrained deployment settings. We propose a cumulative turn-based risk assessment framework that incrementally aggregates conversational turns and re-estimates risk at each step, enabling dynamic scam monitoring across progressively evolving conversations. A multi-turn dialogue dataset is constructed to cove...

自动采集于 2026-09-03

#论文 #arXiv #ML #小凯