Dual-Personalizing Adapter for Federated Foundation Models

TL;DR

This work addresses the practical problem of test-time distribution shifts in federated foundation models. It introduces FedDPA, a dual-adapter approach that learns a global adapter for broad, test-time task knowledge and a local adapter for client-specific personalization, with an instance-wise dynamic weighting scheme to fuse their outputs during inference. Empirical results on two federated NLP benchmarks show state-of-the-art personalization and improved test-time adaptation, with the global adapter contributing to robust generalization and the local adapter enabling tailored performance. The approach underscores the viability of lightweight, privacy-preserving personalization in large-scale foundation models and suggests directions for broader FL frameworks and theoretical analysis.

Abstract

Recently, foundation models, particularly large language models (LLMs), have demonstrated an impressive ability to adapt to various tasks by fine-tuning diverse instruction data. Notably, federated foundation models (FedFM) emerge as a privacy preservation method to fine-tune models collaboratively under federated learning (FL) settings by leveraging many distributed datasets with non-IID data. To alleviate communication and computation overhead, parameter-efficient methods are introduced for efficiency, and some research adapted personalization methods to FedFM for better user preferences alignment. However, a critical gap in existing research is the neglect of test-time distribution shifts in real-world applications, and conventional methods for test-time distribution shifts in personalized FL are less effective for FedFM due to their failure to adapt to complex distribution shift scenarios and the requirement to train all parameters. To bridge this gap, we refine the setting in FedFM, termed test-time personalization, which aims to learn personalized federated foundation models on clients while effectively handling test-time distribution shifts simultaneously. To address challenges in this setting, we explore a simple yet effective solution, a Federated Dual-Personalizing Adapter (FedDPA) architecture. By co-working with a foundation model, a global adapter and a local adapter jointly tackle the test-time distribution shifts and client-specific personalization. Additionally, we introduce an instance-wise dynamic weighting mechanism that dynamically integrates the global and local adapters for each test instance during inference, facilitating effective test-time personalization. The effectiveness of the proposed method has been evaluated on benchmark datasets across different NLP tasks.

Figures (3)

• Figure 1: The overall framework of FedDPA. Each client contains a frozen LLM, a trainable global adapter (LoRA) and a trainable local adapter (LoRA) with a specific task, where the global adapter (LoRA) is for test-time tasks and the local adapter (LoRA) is for personalization. During the training, only the parameters of the global adapter (LoRA) are transmitted to the server for aggregation.
• Figure 2: Frameworks of two personalized methods for local adapter (LoRA) are shown on the left, with their overall learning processes on the right.
• Figure 3: Average accuracy varies as different instance numbers. TTP represents test-time personalization.