An Explainable Agentic System for Detection of Conversational Scams with Summary-Based Memory
The paper proposes an explainable agentic system for detecting conversational scams and introduces a multi-category benchmark dataset, ConScamBench-278. On isolated messages, the single-message detector achieves 100% phishing recall; on the LoveFraud02 corpus, the conversation-level detector identifies all 83 scam conversations; on ConScamBench-278, it achieves 97.8% accuracy. Two user studies (N=100 and N=45) show that participants often feel uncertain about suspicious conversations.
The system achieves conversation-level scam detection via a summary-based memory mechanism, performing excellently on benchmarks. However, user studies reveal uncertainty in real-world scenarios, indicating the need for further optimization of explainability and user trust.
Demand for conversational scam detection is growing. Existing methods mostly focus on single messages; this work fills the gap in multi-turn conversation detection and may drive security products toward agentic evolution.
This technology can be integrated into instant messaging, social platforms, or financial applications to provide real-time scam alerts, reducing the risk of user financial loss.
Attention should be paid to community adoption of the ConScamBench-278 benchmark, as well as the false positive rate and user acceptance of the system in real deployments.