Financial scams increasingly target older adults via text or voice channels such as email, SMS and phone calls. The scams unfold over multiple conversational turns: they start with impersonation or casual contact, then build trust and urgency, and finally request sensitive information or money transfers. Because risk signals appear incrementally across turns, detection models must continuously update risk estimates under resource‑constrained deployment. We introduce a cumulative turn‑based risk assessment framework that aggregates all observed turns and re‑estimates risk after each turn, enabling dynamic monitoring of evolving dialogues. A multi‑turn dialogue dataset is built covering investment, charity and tech‑support scams; each dialogue contains 2‑8 turns and is annotated at every cumulative stage with a qualitative risk level, a continuous risk score, an explanatory rationale and a safety recommendation. Four small language models (Phi‑4, LLaMA‑3.2, DeepSeek‑R1, and Qwen3) are fine‑tuned within a unified training pipeline. The fine‑tuned models capture fraud‑related linguistic cues and cross‑turn escalation patterns while remaining compact enough for mobile or edge deployment. Results show that Phi‑4 and LLaMA‑3.2 achieve stronger turn‑aware risk estimation relative to their parameter size. These findings suggest that structured cumulative modeling can support incremental scam risk assessment in deployment‑oriented settings and highlight the potential of compact LLMs for privacy‑preserving, on‑device fraud protection.
Review