Personalized Mobile Edge Generation: A Stable Personalized Training Approach via Scaling Connection

Mobile Edge Generation (MEG) is presented as a distributed framework in which an identical diffusion model (DM) is deployed on both an edge server (ES) and user equipment (UE). In MEG, most computations and generation steps of UEs are offloaded to the ES. However, heterogeneous user preferences cannot be captured by a uniform DM. To address this, a Personalized Mobile Edge Generation (P-MEG) framework is proposed, where a lightweight personalized U-Net is trained on the UE in collaboration with the pre-trained DM from the ES. During inference, pre-trained ES features are fused with UE features through scaling coefficients that encode user-specific preferences. The training stability of P-MEG and the robustness of feature fusion under noisy wireless channels are theoretically investigated, where bounds are derived on forward and backward feature oscillations, backpropagation gradients, and feature fusion errors in the presence of additive white Gaussian noise (AWGN) noise. These bounds are shown to depend on the fusion scale, and robustness under AWGN follows the same dependence. A multi-U-Net training model with AWGN perturbations is introduced to emulate over-the-air training. Inspired by these insights, a constant scaling connection (CSC) method is proposed to stabilize training by exponentially scaling the fusion coefficients, and a random mask training (RMT) strategy is introduced to reduce computational requirements by adjusting transmission ratios of personalized features. Experimental evaluations on MNIST, EMNIST and PACS demonstrate that: 1) P-MEG enables effective personalized image generation, 2) RMT alleviates computational demands with only slight training overhead, and 3) CSC stabilizes feature oscillations under noisy channels, yielding a 1.4-fold acceleration in training.