Privacy Regularization: Joint Privacy-Utility Optimization in Language Models
This addresses privacy risks for users when training models on sensitive data like emails, offering an alternative to differential privacy with reduced utility degradation.
The paper tackled the problem of language models memorizing training data, which raises privacy concerns, by introducing two privacy-preserving regularization methods that jointly optimize utility and privacy, showing advantages like a favorable utility-privacy trade-off and uniform treatment of subgroups.
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.