Differentially Private Model Compression via Selective Pretraining

TL;DR — We propose a framework to optimize pre-training data to improve the efficiency-utility trade-offs in private learning.

Abstract

Suppose we want to train text prediction models in email clients or word processors. These models, which serve billions of predictions per hour, must preserve the privacy of user data and adhere to specific model size constraints to meet memory, inference time requirements, and to reduce inference cost. 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 {\em subset} of the public dataset that is guided by the private dataset is crucial to train small DP language models. On standard benchmarks, models trained with our new framework achieve state-of-the-art performance, improving upon all the baselines from the literature. Besides performance improvements, our framework also shows that with careful pre-training and private fine-tuning, smaller models can match the performance of much larger models that do not have access to private data, highlighting the promise of private learning as a tool for model compression and efficiency.