Full Description
Cybergrooming poses an increasing risk to young internet users, necessitating proactive, AI-driven interventions. This work introduces an adaptive chatbot GUARD that enhances teenagers' resilience against cyber-grooming through dynamic, human-like interactions tailored to individual vulnerability levels.
The system integrates a large language model—fine-tuned on a stage-tagged PJ dataset—with hierarchical deep reinforcement learning to enable flexible transitions across six grooming stages via both reward and context-based signals. A master policy determines the current stage based on conversational cues, while six stage-specific sub-policies optimize responses by leveraging user sentiment and sub-goal achievement.
Performance is evaluated via simulated interactions to ensure the quality of the chatbot prior to deployment. Ultimately, this research provides an AI-driven tool for enhancing online safety education and preventing cybergrooming.
