Overview
Field: Machine Learning Authors: Parviz Ghafariasl, Weimin Fu, Xiaolong Guo Published: 2026-09-03 arXiv: 2509.00004
Summary
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.
The paper proposes 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.
Key Contributions
- Cumulative turn-based framework: risk is re-estimated at each conversational step as evidence accumulates, suited to progressive multi-turn scams.
- Multi-turn dialogue dataset: covers investment, charity, and tech-support scam scenarios; conversations contain 2–8 turns and are annotated at each cumulative stage with qualitative risk levels, continuous risk scores, explanatory rationales, and safety recommendations.
- Small-model evaluation: four small language models—Phi-4, LLaMA-3.2, DeepSeek-R1, and Qwen3—are fine-tuned and benchmarked under a unified training framework.
- Fine-tuned compact models capture fraud-related linguistic cues and cross-turn escalation patterns while remaining suitable for mobile and resource-constrained deployments.
- Among the evaluated models, Phi-4 and LLaMA-3.2 deliver the strongest turn-aware risk estimation performance relative to their parameter size.
- Structured cumulative modeling can support deployment-oriented incremental scam risk assessment, highlighting the potential of compact language models for privacy-aware, on-device fraud protection.