隐私保护的低秩适应性潜态扩散模型

原文约400字，阅读约需1分钟。发表于：。

Low-rank adaptation is used to adapt latent diffusion models, but it is vulnerable to membership inference attacks; therefore, a privacy-preserving solution called Stable PrivateLoRA is proposed to mitigate this issue and effectively defend against MI attacks while generating high-quality images.

该研究提出了一种新的LLM服务范例，通过在边缘设备上进行隐私敏感计算并在云端共享计算，实现了数据本地性。核心创新PrivateLoRA通过利用低秩性质实现了超过95%的通信减少，同时提供了与LoRA相媲美的调优性能。这是文献中第一个高效且保护隐私的LLM解决方案。