This research addresses the challenge of moderating culturally-specific offensive memes by creating a specialized system for Singapore's diverse linguistic landscape.

Key Innovations:

• Created a dataset of 112K memes labeled by GPT-4V specifically for Singapore context
• Fine-tuned a vision-language model to detect offensive content with cultural awareness
• Developed capability to process multimodal content combining local languages and imagery
• Enhanced security through improved detection of harmful content in culturally nuanced formats

The security implications are significant: traditional content moderation fails with multimodal content in diverse cultural settings, creating blind spots for harmful material. This approach demonstrates how AI can be adapted to protect specific communities with unique cultural contexts.

Detecting Offensive Memes with Social Biases in Singapore Context Using Multimodal Large Language Models