Selective Pre-training for Private Fine-tuning

TL;DR

This paper introduces a pre-training approach on a dataset subset guided by private data, enabling small, efficient, and private language models to achieve state-of-the-art performance while maintaining privacy.

Contribution

It proposes a novel pre-training method on a dataset subset guided by private data, improving small model performance with differential privacy.

Findings

Small models trained with the new framework outperform existing private models.

Careful pre-training allows small models to match larger non-private models.

The approach enhances model efficiency and privacy preservation.

Abstract

Text prediction models, when used in applications like email clients or word processors, must protect user data privacy and adhere to model size constraints. These constraints are crucial to meet memory and inference time requirements, as well as to reduce inference costs. Building small, fast, and private domain-specific language models is a thriving area of research. In this work, we show that a careful pre-training on a \emph{subset} of the public dataset that is guided by the private dataset is crucial to train small language models with differential privacy. On standard benchmarks, small models trained with our new framework achieve state-of-the-art performance. In addition to performance improvements, our results demonstrate that smaller models, through careful pre-training and private fine-tuning, can match the performance of much larger models that do not have access to private data. This underscores the potential of private learning for model compression and enhanced efficiency.