Small LLMs Detect Elder Financial Scams Incrementally

Financial scams targeting older adults are increasingly sophisticated, often unfolding over multiple conversational turns via text or voice. These scams typically begin with casual contact, build trust, create urgency, and eventually lead to requests for sensitive information or financial transfers. Detecting such scams effectively requires a system that can continuously update risk estimates as the conversation progresses, especially in resource-limited environments. The paper introduces a cumulative turn-based risk assessment framework designed to aggregate conversational turns incrementally and re-estimate risk at each step. This allows for dynamic monitoring of evolving scam conversations. To facilitate this, a multi-turn dialogue dataset was created, covering investment, charity, and tech support scams, with each turn annotated for qualitative risk, a continuous risk score, rationale, and safety recommendations. Four small language models (Phi-4, LLaMA-3.2, DeepSeek-R1, and Qwen3) were fine-tuned and evaluated within this framework. The results indicate that these fine-tuned SLMs are capable of identifying fraud-related linguistic cues and cross-turn escalation patterns. Notably, Phi-4 and LLaMA-3.2 demonstrated strong turn-aware risk estimation performance relative to their size, highlighting the potential of compact language models for privacy-aware and on-device fraud protection in deployment-oriented settings.