Split Federated Learning for AIGC with Resource Cognition in Collaborative Mobile Edge Computing

The era of artificial intelligence generated content (AIGC) has emerged to offer the benefits of personalized content to meet the needs of different users. A majority of generative AI models require computationally intensive training and/or finetuning with large datasets, which make resource-constrained devices, such as mobile edge devices (MEDs) to seldom participate in the training process. In addition, MEDs are not willing to share their private data with a central cloud server to train the models. Distributed machine learning (ML) approaches, such as federated learning (FL) and split learning (SL), offer some benefits in achieving data privacy, as well as model convergence for balanced data distributions. However, in practical settings where the datasets of MEDs are imbalanced, both FL and SL suffer poor performance. This motivates us to propose collaborative split federated learning (CSFL) for AIGC. Our CSFL approach supports resource cognition in which MEDs participate based on their available resources. We demonstrate that CFSL achieves good performance in terms of model convergence speed and training latency, while preserving data privacy when training a denoising diffusion probabilistic model for AIGC.