#inference-attacks

Model Theft via API: How Extraction Attacks on Closed LLMs Work

Model extraction attacks let adversaries reconstruct proprietary LLMs by querying their APIs at scale. We examine the mechanics, cost-of-extraction