Enabling Differentially Private Federated Learning for Speech Recognition: Benchmarks, Adaptive Optimizers and Gradient Clipping

TL;DR

This paper introduces the first benchmark for federated learning with differential privacy in speech recognition, proposing layer-wise gradient normalization techniques to address gradient heterogeneity in large models, enabling practical privacy-preserving FL.

Contribution

It establishes a novel benchmark for DP-FL in ASR and develops layer-wise gradient normalization methods to improve convergence in large transformer models.

Findings

Achieves strong user-level differential privacy with minimal WER increase.

Demonstrates viability of DP-FL in ASR with large user populations.

Provides broader insights applicable to scalable privacy-preserving FL for large models.

Abstract

While federated learning (FL) and differential privacy (DP) have been extensively studied, their application to automatic speech recognition (ASR) remains largely unexplored due to the challenges in training large transformer models. Specifically, large models further exacerbate issues in FL as they are particularly susceptible to gradient heterogeneity across layers, unlike the relatively uniform gradient behavior observed in shallow models. As a result, prior works struggle to converge with standard optimization techniques, even in the absence of DP mechanisms. To the best of our knowledge, no existing work establishes a competitive, practical recipe for FL with DP in the context of ASR. To address this gap, we establish \textbf{the first benchmark for FL with DP in end-to-end ASR}. Our approach centers on per-layer clipping and layer-wise gradient normalization: theoretical analysis reveals that these techniques together mitigate clipping bias and gradient heterogeneity across layers in deeper models. Consistent with these theoretical insights, our empirical results show that FL with DP is viable under strong privacy guarantees, provided a population of at least several million users. Specifically, we achieve user-level (7.2, $10^{-9}$)-DP (resp. (4.5, $10^{-9}$)-DP) with only a 1.3% (resp. 4.6%) absolute drop in word error rate when extrapolating to high (resp. low) population scales for FL with DP in ASR. Although our experiments focus on ASR, the underlying principles we uncover - particularly those concerning gradient heterogeneity and layer-wise gradient normalization - offer broader guidance for designing scalable, privacy-preserving FL algorithms for large models across domains. Code of all experiments and benchmarks is available at https://github.com/apple/ml-pfl4asr.