PrivateLoRA For Efficient Privacy Preserving LLM
Yiming Wang, Yu Lin, Xiaodong Zeng, Guannan Zhang
TL;DR
The paper addresses the privacy-versus-efficiency challenge in LLM services by proposing an edge-cloud division of labor. It introduces PrivateLoRA, a method that transmits only activations and exploits low-rank residual activations to minimize communication. It reports substantial gains: over 95% communication reduction, throughput >300% of device-only on 7B models, and ~80% of an A100 GPU for 33B models on 5G networks, with personalization comparable to LoRA. This approach democratizes access to privacy-preserving, high-performance generative AI for edge devices and represents a novel contribution as an efficient privacy-preserving LLM solution.
Abstract
End users face a choice between privacy and efficiency in current Large Language Model (LLM) service paradigms. In cloud-based paradigms, users are forced to compromise data locality for generation quality and processing speed. Conversely, edge device paradigms maintain data locality but fail to deliver satisfactory performance. In this work, we propose a novel LLM service paradigm that distributes privacy-sensitive computation on edge devices and shared computation in the cloud. Only activations are transmitted between the central cloud and edge devices to ensure data locality. Our core innovation, PrivateLoRA, addresses the challenging communication overhead by exploiting the low rank of residual activations, achieving over 95% communication reduction. Consequently, PrivateLoRA effectively maintains data locality and is extremely resource efficient. Under standard 5G networks, PrivateLoRA achieves throughput over 300% of device-only solutions for 7B models and over 80% of an A100 GPU for 33B models. PrivateLoRA also provides tuning performance comparable to LoRA for advanced personalization. Our approach democratizes access to state-of-the-art generative AI for edge devices, paving the way for more tailored LLM experiences for the general public. To our knowledge, our proposed framework is the first efficient and privacy-preserving LLM solution in the literature.