请问HN:端到端加密的LLM聊天(开放模型和闭合模型)

1作者: 5F7bGnd6fWJ66xN24 天前原帖
我正在探索一个软件层——类似于公钥/私钥加密——使用户能够与大型语言模型(LLM)进行对话,同时保证提示和响应对所有中介(包括模型托管方)都是不可读的。(我指的是加密学意义上的“密码”。) 两个案例: 1. 开放权重模型:确保操作员仍然无法读取提示/响应。 2. 闭合托管模型:真正的端到端加密(E2EE),以至于提供者也无法检查内容。 我们可以讨论的话题: - 最佳的短期路径:带有证明的可信执行环境(TEE)、同态加密(FHE/HE)、多方计算(MPC)/分离推理、用于检索的私密信息检索(PIR)、差分隐私或混合方案? - 如何处理前向保密的密钥交换/轮换? - 实际性能/准确性限制(例如,非线性、键值缓存、流式处理)? - 最小可行架构和现实威胁模型? - 你能推荐的任何先前的研究或团队? 如果你有兴趣与我合作,请私信我。
查看原文
I’m exploring a software layer—analogous to public&#x2F;private-key crypto—so a user can converse with an LLM where prompts and responses remain unreadable to all intermediaries, including the model host. (I mean “cipher” in the cryptographic sense.)<p>Two cases: Open-weights model: ensure the operator still can’t read prompts&#x2F;responses. Closed, hosted model: true E2EE so even the provider can’t inspect content.<p>Topics we can discuss: Best near-term path: TEEs with attestation, FHE&#x2F;HE, MPC&#x2F;split inference, PIR for retrieval, differential privacy, or hybrids? How to handle key exchange&#x2F;rotation for forward secrecy? Practical performance&#x2F;accuracy limits (e.g., non-linearities, KV-cache, streaming)? Minimal viable architecture and realistic threat model? Any prior art or teams you’d point me to?<p>Please DM if you are interested in working with me.