Differentially Private Parameter-Efficient Fine-tuning for Large ASR Models

TL;DR

This study explores DP parameter-efficient fine-tuning as a way to mitigate privacy risks with smaller computation and performance costs for ASR models.

Abstract

Large ASR models can inadvertently leak sensitive information, which can be mitigated by formal privacy measures like differential privacy (DP). However, traditional DP training is computationally expensive, and can hurt model performance. Our study explores DP parameter-efficient fine-tuning as a way to mitigate privacy risks with smaller computation and performance costs for ASR models. Through extensive experimentation and progressive optimization, we achieve 4.6%/8.1% word error rate on LibriSpeech clean/other test-sets, setting a new performance benchmark while maintaining (10, 3.52e-6)-DP in fine-tuning a large ASR model with over 600M parameters.

Figures (5)

• Figure 1: ASR model training workflow.
• Figure 2: PEFT methods in ASR models.
• Figure 3: WER_other for different placements of LoRA modules.
• Figure 4: Impact of $\sigma$ in Gaussian initialization on DP-LoRA/DP-RP with $(10, 3.52\mathrm{e}{-6})$-DP.
• Figure 5: Comparison results of WER_clean and WER_other for DP FT/PEFT methods for the same number of fine-tuning steps.

Theorems & Definitions (1)

• Definition 1: Differential privacy dwork2006calibrating