Key insights
5 slides
The solution
A new multimodal deep learning framework achieves 96.72% accuracy in detecting harmful content.
This framework identifies violent, racist, and harassing content within child-friendly online games.
1 / 5
- 01A new multimodal deep learning framework achieves 96.72% accuracy in detecting violent, racist, and harassing content within child-friendly online games.See it on page 1
- 02The system utilizes a hybrid late fusion approach that combines 3D convolutional neural networks and attention mechanisms for visual analysis with BERT and BiLSTM models for text processing.See it on page 21
- 03The model was trained on a curated dataset of 10,702 Indonesian-language game chat interactions to identify prohibited behaviors.See it on page 5
- 04Designed for server-side deployment, the architecture performs real-time monitoring by analyzing five-minute video segments and triggering alerts for parents.See it on page 11
- 05The current implementation relies on deterministic rule-based logic combined with probability weighting to classify user interactions.See it on page 9
- 06Future development plans include expanding language support beyond Indonesian, integrating auditory analysis, and improving detection of evolving online slang.See it on page 17
Ask AI about this article
Summary
The rapid proliferation of online gaming necessitates robust safety mechanisms to protect younger audiences from exposure to harmful content. To address this, a multimodal deep learning framework has been developed to identify visual and verbal violence within child-friendly gaming environments. By integrating visual analysis—utilizing 3D convolutional neural networks, bidirectional long short-term memory networks, and attention mechanisms—with verbal analysis powered by BERT and BiLSTM, the system provides a comprehensive monitoring solution. This architecture employs a hybrid late fusion approach, combining deterministic rule-based logic with probability weighting to classify interactions into neutral, violent, racist, or harassment categories.
The system demonstrates high efficacy, achieving an overall classification accuracy of 96.72%. This performance is supported by a curated dataset of 10,702 Indonesian-language game chats, which serves as a critical resource for training automated early warning systems. Designed to operate on server-side infrastructure, the model processes five-minute video segments to detect prohibited behaviors in real-time, subsequently alerting parents to potential risks. This proactive monitoring capability represents a significant advancement in digital child safety, moving beyond static filtering to dynamic, context-aware content moderation.
While the current implementation provides a strong foundation for monitoring online gaming, the scope of the research is primarily focused on the Indonesian linguistic context and visual-textual integration. Future development efforts are directed toward expanding language support, incorporating auditory modalities, and enhancing the model’s robustness against the rapidly evolving nature of online slang. By grounding these advancements in established deep learning architectures, the system offers a scalable framework for mitigating toxicity and ensuring safer digital experiences for children across diverse gaming platforms.
