The Risk of Federated Learning to Skew Fine-Tuning Features and Underperform Out-of-Distribution Robustness

TL;DR

Federated fine-tuning on privacy-sensitive, domain-specific data can skew feature representations and degrade out-of-distribution robustness. The authors introduce three robustness indicators—$SVE$, $LSVR$, and $GDA$—and a general noisy projection-based algorithm (GNP) that transfers robustness from the pre-trained model to the fine-tuned model while augmenting capacity via Gaussian noise. Through experiments on multiple robust NLP datasets and several PEFT methods, they show that data heterogeneity and choice of PEFT can undermine OOD robustness, and that GNP consistently improves robustness without sacrificing in-distribution performance. The proposed framework offers a practical, general approach to preserving OOD robustness in federated, parameter-efficient fine-tuning regimes, with broad relevance to real-world applications in NLP under privacy constraints.

Abstract

To tackle the scarcity and privacy issues associated with domain-specific datasets, the integration of federated learning in conjunction with fine-tuning has emerged as a practical solution. However, our findings reveal that federated learning has the risk of skewing fine-tuning features and compromising the out-of-distribution robustness of the model. By introducing three robustness indicators and conducting experiments across diverse robust datasets, we elucidate these phenomena by scrutinizing the diversity, transferability, and deviation within the model feature space. To mitigate the negative impact of federated learning on model robustness, we introduce GNP, a \underline{G}eneral \underline{N}oisy \underline{P}rojection-based robust algorithm, ensuring no deterioration of accuracy on the target distribution. Specifically, the key strategy for enhancing model robustness entails the transfer of robustness from the pre-trained model to the fine-tuned model, coupled with adding a small amount of Gaussian noise to augment the representative capacity of the model. Comprehensive experimental results demonstrate that our approach markedly enhances the robustness across diverse scenarios, encompassing various parameter-efficient fine-tuning methods and confronting different levels of data heterogeneity.

Figures (10)

• Figure 1: Two-Stage deployment process of large language models.
• Figure 2: The probability distributions of client data labels under different Dirichlet parameter $\alpha$. Smaller $\alpha$ indicates a higher degree of data heterogeneity.
• Figure 3: Comparative heatmap of accuracy under different $\alpha$ on Amazon, Dynasent, Semeval, and SST datasets. The vertical axis represents four different fine-tuning methods: Full fine-tuning (FT), LoRA (LR), prefix tuning (PF), adapter tuning (AP), and BitFit (BF). The horizontal axis represents various $\alpha$ values, with increasing data heterogeneity from left to right. The heatmap uses red for datasets with the ID datasets and blue for robust datasets, where darker colors indicate higher accuracy.
• Figure 4: Evolution of three robust indicators with varying data heterogeneity.
• Figure 5: A boxplot comparison illustrates the accuracy difference between the full fine-tuning method and four parameter-efficient fine-tuning methods. Significance differences among the groups are indicated at the top of each subset, with larger values denoting lower differences. The blue, yellow, and red lines represent accuracy at different levels of data heterogeneity.

Theorems & Definitions (3)

• Definition 1: Singular Value Entropy, SVE
• Definition 2: Largest Singular Value Ratio, LSVR
• Definition 3: Gradient Deviation Angle, GDA