Model Extraction Attacks and Defenses for LLMs

Capture Summary

Survey of LLM-specific model extraction attacks and defenses. Search result categorizes attacks into functionality extraction, training data extraction, and prompt-targeted attacks, with defenses for model protection, data privacy, and prompt protection.

Relevance

• Fills a current gap in the vault's model extraction coverage.
• Useful for research on hosted API protection, watermarking, behavioral detection, and query auditing.

Collection Notes

Collected as model extraction survey.