Chenhang Christopher Zhang¹ and Andrew Park², ¹Concord Academy, USA, ²California State Polytechnic University, USA

With the growing risks for children online, this project introduces a Chrome extension that leverages AI-driven text and image classification to filter harmful content in real time. The system employs BERT for text classification and YOLOv8 for image analysis, dynamically blocking inappropriate material while allowing safe content to pass through. Through experimental evaluation, we identified classification weaknesses, prompting the removal of the monetary and social categories due to persistent misclassification. Iterative dataset refinement and model retraining led to significant performance improvements. Performance testing revealed that GPU acceleration is essential for real-time filtering, as CPU-based deployment resulted in substantial delays. Future work will focus on further dataset refinement, model optimization, and multimodal AI integration to enhance efficiency and accuracy. The results demonstrate the viability of AI-powered real-time filtering, offering a customizable and adaptive approach to online safety. This study lays the groundwork for future advancements in automated content moderation, contributing to a safer digital environment for children.

Online Safety, Child Internet Use, AI Content Filtering, Parental Controls, Harmful Content Detection, Web Filtering, Nonprofit Technology