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

Flocks of Stochastic Parrots: Differentially Private Prompt Learning for Large Language Models

About

Large language models (LLMs) are excellent in-context learners. However, the sensitivity of data contained in prompts raises privacy concerns. Our work first shows that these concerns are valid: we instantiate a simple but highly effective membership inference attack against the data used to prompt LLMs. To address this vulnerability, one could forego prompting and resort to fine-tuning LLMs with known algorithms for private gradient descent. However, this comes at the expense of the practicality and efficiency offered by prompting. Therefore, we propose to privately learn to prompt. We first show that soft prompts can be obtained privately through gradient descent on downstream data. However, this is not the case for discrete prompts. Thus, we orchestrate a noisy vote among an ensemble of LLMs presented with different prompts, i.e., a flock of stochastic parrots. The vote privately transfers the flock's knowledge into a single public prompt. We show that LLMs prompted with our private algorithms closely match the non-private baselines. For example, using GPT3 as the base model, we achieve a downstream accuracy of 92.7% on the sst2 dataset with ($\epsilon=0.147, \delta=10^{-6}$)-differential privacy vs. 95.2% for the non-private baseline. Through our experiments, we also show that our prompt-based approach is easily deployed with existing commercial APIs.

Haonan Duan, Adam Dziedzic, Nicolas Papernot, Franziska Boenisch• 2023

Related benchmarks

TaskDatasetResultRank
Text ClassificationAG News (test)
Accuracy74.6
293
Text ClassificationSST-2 (test)
Accuracy95.7
185
Text ClassificationTREC (test)
Accuracy79.3
122
Text ClassificationDBPedia (test)--
40
Membership Inference AttackNewsQA
AUC92.9
39
Membership Inference AttackSQuAD
AUC0.679
39
Sentiment ClassificationMPQA
Accuracy92.1
35
Membership Inference AttackSQuADShifts (test)
AUC0.619
26
Sentiment ClassificationSST2
Accuracy95.7
20
Text ClassificationDisaster
Accuracy77.8
17
Showing 10 of 11 rows

Other info

Follow for update