This research provides a comprehensive survey of Membership Inference Attacks (MIAs) on large language and multimodal models, highlighting critical security vulnerabilities.

• MIAs can determine if specific data was used to train a model, creating significant privacy risks
• The survey systematically categorizes attack methodologies across both LLMs and multimodal models
• Researchers identify key vulnerabilities in current model architectures and deployment strategies
• The paper analyzes existing defense mechanisms and their effectiveness against evolving attack vectors

As AI models become increasingly embedded in business workflows, understanding these security implications is essential for responsible implementation and governance of AI systems.

Membership Inference Attacks on Large-Scale Models: A Survey