A Federated Framework for LLM-based Recommendation

TL;DR

Privacy concerns in fine-tuning LLM-based recommender systems motivate Fed4Rec, but directly applying it to LLMs leads to exacerbated client performance imbalance and high resource costs. We propose FELLRec, a federated framework combining a Dynamic Balance Strategy (attention-based dynamic parameter aggregation and Curriculum Heating learning speed) with a Flexible Storage Strategy (offloading non-sensitive LLM layers to the server while keeping input/output layers on the client). Empirical results across three datasets and two LLM backends show that FELLRec achieves balanced, privacy-preserving performance that rivals centralized LLM-based methods, while reducing client resource demands and maintaining robustness against attack risk depending on layer-offloading choices. The work advances private, scalable LLM-based recommendations by enabling personalized client updates and resource-aware inference in a federated setting, with practical implications for deploying large models in privacy-sensitive environments.

Abstract

Large Language Models (LLMs) have empowered generative recommendation systems through fine-tuning user behavior data. However, utilizing the user data may pose significant privacy risks, potentially leading to ethical dilemmas and violations of data protection regulations. To address the privacy concerns, Federated Learning for Recommendation (Fed4Rec) has been identified as a promising solution. However, directly applying Fed4Rec in the LLM context introduces two challenges: 1) exacerbated client performance imbalance, which ultimately impacts the system's long-term effectiveness, and 2) substantial client resource costs, posing a high demand for clients' both computational and storage capability to locally train and infer LLMs. To tackle these challenges, we propose a federated framework for LLM-based recommendation (shorted as FELLRec). Generally, FELLRec designs two key strategies. 1) Dynamic balance strategy, which designs dynamic parameter aggregation and learning speed for different clients, aiming to ensure balanced performance across clients. 2) Flexible storage strategy, which selectively retains certain sensitive LLM layers on the client side, while offloading other layers to the server, aiming to preserve privacy while saving resources. Experiment results show that FELLRec can achieve a more balanced client performance and improved overall performance in a computational and storage-efficient way while safeguarding user privacy well.

Figures (7)

• Figure 1: (a) illustrates the exacerbated client performance imbalance when applying a classical client-server method (FedAvg mcmahan2017communication) to LLM-based recommender models (BIGRec) compared with traditional recommender models (MF). (b) shows the convergence rate of two selected clients when applying FedAvg to LLM-based and traditional models. The observations are on Games.
• Figure 2: FELLRec Structure. The left part is the flexible storage strategy which offloads non-sensitive LLM layers to the server to save resources. The right part is the dynamic balance strategy which ensures relatively balanced performance across clients.
• Figure 3: (a) and (b) shows the similarity between input embeddings and predicted input embeddings according to extracted embeddings of different layers from BIGRec under linear probe attack and MLP probe attack.
• Figure 4: (a) and (b) show the time (s) and memory (MiB) cost for different values of $k$ under the BIGRec-based and RecFormer-based FELLRec, respectively.
• Figure 5: Contributions of dynamic parameter aggregation and learning speed to FELLRec.