DPFormer: Learning Differentially Private Transformer on Long-Tailed Data

Abstract

The Transformer has emerged as a versatile and effective architecture with broad applications. However, it still remains an open problem how to efficiently train a Transformer model of high utility with differential privacy guarantees. In this paper, we identify two key challenges in learning differentially private Transformers, i.e., heavy computation overhead due to per-sample gradient clipping and unintentional attention distraction within the attention mechanism. In response, we propose DPFormer, equipped with Phantom Clipping and Re-Attention Mechanism, to address these challenges. Our theoretical analysis shows that DPFormer can reduce computational costs during gradient clipping and effectively mitigate attention distraction (which could obstruct the training process and lead to a significant performance drop, especially in the presence of long-tailed data). Such analysis is further corroborated by empirical results on two real-world datasets, demonstrating the efficiency and effectiveness of the proposed DPFormer.