ChartAttack: Testing the Vulnerability of LLMs to Malicious Prompting in Chart Generation
Researchers have developed a framework called ChartAttack to test the vulnerability of multimodal large language models (MLLMs) to malicious prompting in chart generation. This framework injects misleading elements into chart designs, which can lead to incorrect interpretations by both AI and humans. Experiments showed that ChartAttack significantly reduced MLLM accuracy on chart question-answering tasks, highlighting the need for enhanced robustness and security in MLLM-based chart generation systems. AI
IMPACT Highlights critical security risks in AI-driven data visualization, necessitating improved robustness in deployed models.