Phone fraud is a growing concern, affecting people across all demographics and causing substantial financial losses. Despite efforts by government agencies and telecommunication providers to combat this issue, such as raising awareness and helping identify suspicious numbers, challenges persist. There is no Hong Kong local phone scam dataset. The ever-evolving tactics of scammers hinder the collection of original victim dialogues. Current methods rarely capture conversations before deception occurs, impeding the development of early intervention strategies. Reports from victims often lack critical dialogue and prosody details, limiting the ability to analyse fraud scenarios and emotional manipulation, thereby reducing the effectiveness of prevention measures.
This project aims to develop an AI-powered honeypot system to address these challenges. The system will engage scammers in phone conversations by generating quick, realistic responses with tone variations, prolonging interactions to waste their time and resources. By capturing original dialogues, the system will conduct semantic analysis to detect suspicious fraud scenarios, enabling early identification of scam numbers and tracking emerging fraud trends for preventive measures. Additionally, it will perform acoustic analysis to evaluate the effectiveness of AI-generated dialogues, using reinforcement learning to enhance system dialogue generation performance over time.