Logits of API-Protected LLMs Leak Proprietary Information

arXiv:2403.09539v1 Announce Type: cross
Abstract: The commercialization of large language models (LLMs) has led to the common practice of high-level API-only access to proprietary models. In this work, we show that even with a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1,000 for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most …

abstract api architecture arxiv cs.ai cs.cl cs.cr cs.lg information language language models large language large language models leak learn llms practice proprietary proprietary models public show type work