Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fine-Tuning Language Models with Differential Privacy through Adaptive Noise Allocation

Xianzhi Li, Ran Zmigrod, Zhiqiang Ma, Xiaomo Liu, Xiaodan Zhu

TL;DR

ANADP, a novel algorithm that adaptively allocates additive noise based on the importance of model parameters, is proposed, demonstrating that ANADP narrows the performance gap between regular fine-tuning and traditional DP fine-tuning on a series of datasets while maintaining the required privacy constraints.

Abstract

Language models are capable of memorizing detailed patterns and information, leading to a double-edged effect: they achieve impressive modeling performance on downstream tasks with the stored knowledge but also raise significant privacy concerns. Traditional differential privacy based training approaches offer robust safeguards by employing a uniform noise distribution across all parameters. However, this overlooks the distinct sensitivities and contributions of individual parameters in privacy protection and often results in suboptimal models. To address these limitations, we propose ANADP, a novel algorithm that adaptively allocates additive noise based on the importance of model parameters. We demonstrate that ANADP narrows the performance gap between regular fine-tuning and traditional DP fine-tuning on a series of datasets while maintaining the required privacy constraints.

Fine-Tuning Language Models with Differential Privacy through Adaptive Noise Allocation

TL;DR

ANADP, a novel algorithm that adaptively allocates additive noise based on the importance of model parameters, is proposed, demonstrating that ANADP narrows the performance gap between regular fine-tuning and traditional DP fine-tuning on a series of datasets while maintaining the required privacy constraints.

Abstract

Language models are capable of memorizing detailed patterns and information, leading to a double-edged effect: they achieve impressive modeling performance on downstream tasks with the stored knowledge but also raise significant privacy concerns. Traditional differential privacy based training approaches offer robust safeguards by employing a uniform noise distribution across all parameters. However, this overlooks the distinct sensitivities and contributions of individual parameters in privacy protection and often results in suboptimal models. To address these limitations, we propose ANADP, a novel algorithm that adaptively allocates additive noise based on the importance of model parameters. We demonstrate that ANADP narrows the performance gap between regular fine-tuning and traditional DP fine-tuning on a series of datasets while maintaining the required privacy constraints.
Paper Structure (16 sections, 7 equations, 3 figures, 2 tables, 1 algorithm)

This paper contains 16 sections, 7 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Differential Privacy.
  4. Fine-Tuning and PEFT.
  5. The AnaDP Model
  6. Parameter Importance.
  7. Importance Normalization.
  8. Experiments
  9. Experimental Setup
  10. Accuracy Results
  11. Exposure Risks Results
  12. AnaDP Noise Distribution
  13. Conclusion
  14. Limitations
  15. Appendix A
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Privacy leakage risks for telephone number and person name using DP, AnaDP and non-DP full fine-tuning.
  • Figure 2: Distribution of noise multipliers during the training of AnaDP on the SST-2 dataset. The X-axis represents the 12 layers of the Roberta-base model, while the Y-axis denotes the PEFT weights. The color gradient indicates the varying amounts of noise applied.
  • Figure 3: Performance of AnaDP under different privacy budget on QNLI dataset.