# DP-NMT: Scalable Differentially-Private Machine Translation

**Authors:** Timour Igamberdiev, Doan Nam Long Vu, Felix Künnecke, Zhuo Yu, Jannik Holmer, Ivan Habernal  
**arXiv:** [2311.14465](https://arxiv.org/abs/2311.14465)  
**Date:** 2023-11-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.14465v1/

## Key points

- Research gap in privacy-preserving neural machine translation (NMT) models
- Lack of clarity in implementing differentially private stochastic gradient descent (DP-SGD) in existing models
- Introduction of DP-NMT, an open-source framework for privacy-preserving NMT with DP-SGD
- Bringing together various models, datasets, and evaluation metrics in a systematic software package
- Importance of clarifying implementation details specific to privacy settings
- Need to understand differences between random shuffling and Poisson sampling in terms of privacy guarantees
- No research currently incorporates DP-SGD into an NMT system
- Framework aims to transparently and intuitively implement the DP-SGD algorithm
- Conducted experiments on datasets from general and privacy-related domains
- Framework made publicly available, welcomes feedback from the community

## Summary

The paper discusses the research gap in the development of privacy-preserving neural machine translation (NMT) models and the lack of clarity in implementing differentially private stochastic gradient descent (DP-SGD) in existing models. To address these issues, the authors introduce DP-NMT, an open-source framework for privacy-preserving NMT with DP-SGD. The framework brings together various models, datasets, and evaluation metrics in a systematic software package to provide a platform for researchers to advance the development of privacy-preserving NMT systems. The authors emphasize the importance of clarifying implementation details specific to privacy settings as they may have significant implications for privacy amplification gains. They highlight the need to understand how random shuffling and Poisson sampling differ in terms of privacy guarantees. While there have been studies on NMT with federated learning and differential privacy included in parameter aggregation, there is currently no research that incorporates DP-SGD into an NMT system. The authors aim to fill this gap by providing a comprehensive framework that transparently and intuitively implements the DP-SGD algorithm. To demonstrate the effectiveness of their framework, they conduct experiments on datasets from both general and privacy-related domains. They make their framework publicly available and welcome feedback from the community. In conclusion, this paper introduces DP-NMT as an open-source framework for scalable differentially private machine translation. It addresses the research gap in privacy preserving NMT models and provides a platform for researchers to advance their development while ensuring transparency and reproducibility.