Selective Differential Privacy for Language Modeling (2108.12944v3)

Abstract: With the increasing applications of LLMs, it has become crucial to protect these models from leaking private information. Previous work has attempted to tackle this challenge by training RNN-based LLMs with differential privacy guarantees. However, applying classical differential privacy to LLMs leads to poor model performance as the underlying privacy notion is over-pessimistic and provides undifferentiated protection for all tokens in the data. Given that the private information in natural language is sparse (for example, the bulk of an email might not carry personally identifiable information), we propose a new privacy notion, selective differential privacy, to provide rigorous privacy guarantees on the sensitive portion of the data to improve model utility. To realize such a new notion, we develop a corresponding privacy mechanism, Selective-DPSGD, for RNN-based LLMs. Besides LLMing, we also apply the method to a more concrete application--dialog systems. Experiments on both LLMing and dialog system building show that the proposed privacy-preserving mechanism achieves better utilities while remaining safe under various privacy attacks compared to the baselines. The data and code are released at https://github.com/wyshi/lm_privacy to facilitate future research .