Privacy Regularization: Joint Privacy-Utility Optimization in Language Models

Abstract

Neural language models are known to have a high capacity for memorization of training samples. This may have serious privacy implications when training models on user content such as email correspondence. Differential privacy (DP), a popular choice to train models with privacy guarantees, comes with significant costs in terms of utility degradation and disparate impact on subgroups of users. In this work, we introduce two privacy-preserving regularization methods for training language models that enable joint optimization of utility and privacy through (1) the use of a discriminator and (2) the inclusion of a triplet-loss term. We compare our methods with DP through extensive evaluation. We show the advantages of our regularizers with favorable utility-privacy trade-off, faster training with the ability to tap into existing optimization approaches, and ensuring uniform treatment of under-represented subgroups.

NAACL-HLT 2021 Paper

References in corpus (7)

• Deep Learning with Differential Privacy
• The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks
• Privacy in Deep Learning: A Survey
• Differentially Private Learning Needs Better Features (or Much More Data)
• Training Data Leakage Analysis in Language Models
• Enabling Fast Differentially Private SGD via Just-in-Time Compilation and Vectorization
• Training Production Language Models without Memorizing User Data

Cited by in corpus (2)

• Membership Inference on Word Embedding and Beyond
• Benchmarking Differential Privacy and Federated Learning for BERT Models