Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Differentially Private Training of Mixture of Experts Models

Pierre Tholoniat, Huseyin A. Inan, Janardhan Kulkarni, Robert Sim

TL;DR

This paper tackles privacy in training large NLP models by studying Differential Privacy (DP) training for Mixture of Experts (MoE) models, leveraging MoE's computational efficiency (e.g., Switch Transformer) for scalable privacy-preserving training. It identifies three DP-related challenges for MoEs—per-sample gradient computation, per-sample load-balancing, and cross-device gradient clipping with expert parallelism—and proposes practical strategies to address them. In experiments on SST-2 and MNLI with a Switch-Transformer architecture using a privacy budget of $\epsilon=8$ and $\delta=1/N$, and a per-sample clipping norm of 1.0, private MoE fine-tuning achieves 92.0% on SST-2 and 78.7% on MNLI, compared to 94.5% and 85.4% non-private baselines. This work demonstrates feasibility and offers a blueprint for scaling privacy-preserving MoEs to larger models and broader NLP tasks.

Abstract

This position paper investigates the integration of Differential Privacy (DP) in the training of Mixture of Experts (MoE) models within the field of natural language processing. As Large Language Models (LLMs) scale to billions of parameters, leveraging expansive datasets, they exhibit enhanced linguistic capabilities and emergent abilities. However, this growth raises significant computational and privacy concerns. Our study addresses these issues by exploring the potential of MoE models, known for their computational efficiency, and the application of DP, a standard for privacy preservation. We present the first known attempt to train MoE models under the constraints of DP, addressing the unique challenges posed by their architecture and the complexities of DP integration. Our initial experimental studies demonstrate that MoE models can be effectively trained with DP, achieving performance that is competitive with their non-private counterparts. This initial study aims to provide valuable insights and ignite further research in the domain of privacy-preserving MoE models, softly laying the groundwork for prospective developments in this evolving field.

Differentially Private Training of Mixture of Experts Models

TL;DR

This paper tackles privacy in training large NLP models by studying Differential Privacy (DP) training for Mixture of Experts (MoE) models, leveraging MoE's computational efficiency (e.g., Switch Transformer) for scalable privacy-preserving training. It identifies three DP-related challenges for MoEs—per-sample gradient computation, per-sample load-balancing, and cross-device gradient clipping with expert parallelism—and proposes practical strategies to address them. In experiments on SST-2 and MNLI with a Switch-Transformer architecture using a privacy budget of and , and a per-sample clipping norm of 1.0, private MoE fine-tuning achieves 92.0% on SST-2 and 78.7% on MNLI, compared to 94.5% and 85.4% non-private baselines. This work demonstrates feasibility and offers a blueprint for scaling privacy-preserving MoEs to larger models and broader NLP tasks.

Abstract

This position paper investigates the integration of Differential Privacy (DP) in the training of Mixture of Experts (MoE) models within the field of natural language processing. As Large Language Models (LLMs) scale to billions of parameters, leveraging expansive datasets, they exhibit enhanced linguistic capabilities and emergent abilities. However, this growth raises significant computational and privacy concerns. Our study addresses these issues by exploring the potential of MoE models, known for their computational efficiency, and the application of DP, a standard for privacy preservation. We present the first known attempt to train MoE models under the constraints of DP, addressing the unique challenges posed by their architecture and the complexities of DP integration. Our initial experimental studies demonstrate that MoE models can be effectively trained with DP, achieving performance that is competitive with their non-private counterparts. This initial study aims to provide valuable insights and ignite further research in the domain of privacy-preserving MoE models, softly laying the groundwork for prospective developments in this evolving field.
Paper Structure (22 sections, 6 equations, 1 figure, 1 table)

This paper contains 22 sections, 6 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Our Contributions
  3. Preliminaries
  4. Differential Privacy
  5. Mixture of Experts
  6. Training MoE models with DP
  7. Per-sample balancing loss
  8. Expert per-sample gradient computation
  9. Per-sample gradients with expert parallelism
  10. Experiments
  11. Model
  12. Datasets
  13. Hyperparameters
  14. Results
  15. Related Work
  16. ...and 7 more sections

Figures (1)

  • Figure 1: Routing with extra batch dimension. Notations are introduced in the "Expert per-sample gradient computation" section.

Theorems & Definitions (1)

  • Definition 1: Differential Privacy (DP) DworkKMMN06