Jinze Li, Yang Zhang, Xin Yang, Jiayi QU, Jinfeng Xu, Shuo Yang, Junhua Ding, Edith Cheuk-Han Ngai. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Sleep health monitoring increasingly relies on wear-able and motion-sensor data, but privacy concerns and data fragmentation hinder large-scale collaboration. In this paper, we propose a Federated Diffusion Framework (FDF) that enables distributed learning of sleep actigraphy data in cloud-edge systems while preserving data privacy. FDF integrates diffusion models with federated optimization and lightweight privacy mechanisms to synthesize realistic sleep-motion data without sharing raw signals. Preliminary experiments on public sleep datasets demonstrate that FDF delivers competitive fidelity and stable downstream utility while reducing communication by approximately 91% and achieving the lowest membership-inference risk among compared baselines. These results highlight the potential of diffusion-based generative modeling for privacy-sensitive healthcare applications and lay the groundwork for scalable, privacy-aware digital twin systems.

Because the dependence, competition, correlation, and other complex interactions among the components and between system and environment, the complex system is difficult to be modeled directly. As an important spacecraft system, satellite needs to meet the high reliability demand. However, the increasingly complex satellite system structure and complex space environment interference make the on-orbit satellite failure inevitable. Modeling and evaluation of satellite conditions can identify potential problems of key system components in time. Firstly, in order to accurately describe the system components and measure the reliability, the correlative damage coefficient is introduced to describe the component degradation process, and the component reliability model of the satellite attitude control system is established. Then, to solve the problem of large state space and changeable condition of system reliability model, the concept of tangent order pair is introduced into the multi-state dynamic fault tree of system. The depth-first search algorithm is used to sort the underlying event components of the fault tree, and the multi-state multi-value decision graph model of different state events is constructed. The minimum cut set of the model is obtained to evaluate the system state. The experimental results show that the proposed method can effectively simulate the system situation under different fault modes, effectively reduce the state space of the system reliability model, and improve the convergence and accuracy of the system fault probability evaluation.