AIGC-Enhanced Federated Learning: Addressing Data Scarcity in Preference-Based Scenarios

Federated learning (FL) is a decentralized machine learning paradigm that enables collaborative model training while preserving data privacy by avoiding sensitive data exfiltration from local storage. Despite its success in various domains, current FL frameworks lack mechanisms to accommodate personalized training objectives, particularly in optimizing performance for specific data classes. Recent advancements in Artificial Intelligence Generated Content (AIGC) present opportunities to address these limitations by supplementing training data for user-preferred classes using generative models. Nonetheless, incorporating AIGC into FL introduces significant challenges, including non-compliant data quality, disorganized data distributions, limited computational resources, and slow data generation speed on edge devices. To address these challenges, we propose AIGC-enhanced Federated Preference Learning (FPL), a novel framework designed to enhance FL performance for user-specified preference classes (PCs). Our approach employs pre-training and fine-tuning of generative models across diverse datasets to improve the quality of synthetic data. Additionally, we optimize FPL efficiency through a client selection strategy that matches tasks involving generated data with suitable clients and a data distribution mechanism that allocates synthetic data to where it is most needed. To further accelerate data supplementation, data augmentation is utilized on local clients. We provide theoretical convergence guarantees for AIGC-enhanced FPL and demonstrate its effectiveness through comprehensive experiments on MNIST and CIFAR-10 datasets, including ablation studies on various data supplementation techniques.