Exploring edge-driven collaborative fine-tuning toward customized AIGC services

With the popularization of artificial intelligence-generated content (AIGC), the explosion of end users is spurring an unprecedented diversity in preferences. To cope with this challenge, the AIGC paradigm is shifting from cloud-driven pre-trained large models to edge-driven personalized customization models. The former are large-scale models with more than 100 billion parameters trained based on massive data, while the latter introduce heterogeneous user preferences or professional knowledge to fine-tune the former. However, such fine-tuning incurs costly resource consumption and privacy disclosure. In this paper, we offer a holistic perspective on the AIGC fine-tuning framework spanning from end user to edge to cloud. Specifically, we first investigate how to deploy an edge-driven collaborative fine-tuning task through federated learning. Then, we discuss a verifiable model consensus protocol and fairness incentive design for edge servers to participate in a collaborative learning task. In addition, a case study focuses on the medical scenario, where we develop a practical X-ray diagnostic demo through collaborative fine-tuning of a multimodal pre-training model, and the diagnostic dialogue and performance results are compared. Finally, potential research directions are identified to advance the edge-driven customized AIGC services.