Privacy Regularization: Joint Privacy-Utility Optimization in LanguageModels

By Fatemehsadat Mireshghallah, Huseyin A. Inan, Marcello Hasegawa, Victor Ruhle, Taylor Berg-Kirkpatrick, Robert Sim. 2021

55 citations recorded by the source corpus.

Neural language models are known to have a high capacity for memorization of training samples. This may have serious privacy im- plications 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 novel 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.

Subfields: Artificial Intelligence, Sociology and Political Science

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Sociology and Political Science

Citations in the corpus

Do Language Models Plagiarize? (2022, background): “, 2021; Leino and Fredrikson, 2020) or extracting artificially injected canaries (Henderson et al., 2018; Mireshghallah et al., 2021; Zanella-Béguelin et al., 2020).”
Submix: Practical Private Prediction for Large-Scale Language Models (2022, background): “Mireshghallah et al. (2021) propose adding a privacy regularizer to language model training to reduce its memorization of sensitive text.”
Mix and Match: Learning-free Controllable Text Generationusing Energy Language Models (2022, background): “…errors associated with the training of language models, but they are also dependent on access to a large amount of attribute-speciﬁc language data which can be impractical in many scenarios and exacerbate privacy concerns (Brown et al., 2022; Mireshghallah et al., 2021; Kandpal et al., 2022).”
You Don’t Know My Favorite Color: Preventing Dialogue Representations from Revealing Speakers’ Private Personas (2022, unknown): “For optimization, we can leave out constant terms and the logarithm (Mireshghallah et al., 2021) to obtain the following loss function: LD(u; θA) = − 1 C C−1∑ k=0 Pr(k|f(u), θA).”
Membership Inference Attacks against Language Models via Neighbourhood Comparison (2023, background): “, 2019) and specific regularization techniques (Mireshghallah et al., 2021; Chen et al., 2022) showing good empirical performance, differ-”
Large Language Models Can Be Strong Differentially Private Learners (2021, background): “Orthogonally, many works considered training language models that satisfy empirical notions of privacy (Xu et al., 2021; Coavoux et al., 2018; Mireshghallah et al., 2021; Melamud & Shivade, 2019).”