🤖 AI Summary
This work investigates whether public pretraining can improve the performance of differentially private (DP) models when substantial distribution shift exists between the public pretraining data and private fine-tuning data. Addressing the challenge that private sensitive data are inaccessible and severe distribution mismatch renders standard DP fine-tuning ineffective, we propose learning low-dimensional shared representations from public data. We establish, for the first time, that even when zero-shot transfer fails completely, these representations significantly boost accuracy in private fine-tuning; theoretical analysis further reveals their sample complexity advantage. Evaluated across three diverse tasks, our method achieves up to a 67% absolute accuracy improvement over DP models trained from scratch. These results empirically validate—and theoretically ground—the efficacy of public representations for practical private learning under extreme distribution shift.
📝 Abstract
Public pretraining is a promising approach to improve differentially private model training. However, recent work has noted that many positive research results studying this paradigm only consider in-distribution tasks, and may not apply to settings where there is distribution shift between the pretraining and finetuning data -- a scenario that is likely when finetuning private tasks due to the sensitive nature of the data. In this work, we show empirically across three tasks that even in settings with large distribution shift, where both zero-shot performance from public data and training from scratch with private data give unusably weak results, public features can in fact improve private training accuracy by up to 67% over private training from scratch. We provide a theoretical explanation for this phenomenon, showing that if the public and private data share a low-dimensional representation, public representations can improve the sample complexity of private training even if it is impossible to learn the private task from the public data alone. Altogether, our results provide evidence that public data can indeed make private training practical in realistic settings of extreme distribution shift.
