Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

DPZero: Private Fine-Tuning of Language Models without Backpropagation

About

The widespread practice of fine-tuning large language models (LLMs) on domain-specific data faces two major challenges in memory and privacy. First, as the size of LLMs continues to grow, the memory demands of gradient-based training methods via backpropagation become prohibitively high. Second, given the tendency of LLMs to memorize training data, it is important to protect potentially sensitive information in the fine-tuning data from being regurgitated. Zeroth-order methods, which rely solely on forward passes, substantially reduce memory consumption during training. However, directly combining them with standard differentially private gradient descent suffers more as model size grows. To bridge this gap, we introduce DPZero, a novel private zeroth-order algorithm with nearly dimension-independent rates. The memory efficiency of DPZero is demonstrated in privately fine-tuning RoBERTa and OPT on several downstream tasks. Our code is available at https://github.com/Liang137/DPZero.

Liang Zhang, Bingcong Li, Kiran Koshy Thekumparampil, Sewoong Oh, Niao He• 2023

Related benchmarks

TaskDatasetResultRank
Image ClassificationCIFAR-10 (test)
Accuracy98.2
882
Image ClassificationCIFAR-100 (test)
Accuracy80.8
295
Question AnsweringSQuAD (test)
F180.1
156
Image ClassificationMNIST (test)
Accuracy80.2
138
Question AnsweringDROP
F1 Score28.4
45
Binary ClassificationSST-2 (test)
Accuracy93.2
32
ClassificationBoolQ (test)
Accuracy66.8
22
Showing 7 of 7 rows

Other info

Follow for update