Sequential federated learning (SFL) enables collaborative model training across clients in a chain manner, providing communication‐efficient benefits over traditional all‐gather parallel federated learning (PFL). However, SFL training often suffers from slow convergence and performance degradation due to nonidentically distributed (non‐IID) data distribution. In motivation experiment, we find that model decoupling by partitioning the model into shared and personalized parameters and using just a few personalized parameters with large gradients can improve SFL training performance. Based on above findings, we propose SeqFedRPC, a novel model decoupling based SFL framework with regularized parameter clustering. We introduce a regularization term to promote parameter sparsification and amplify gradient differences, which aids in gradient‐based parameter clustering. Then, we employ a clustering‐based scheme to adaptively decouple the model parameters into shared and personalized subsets, thereby addressing the challenge of non‐IID data by adapting global knowledge with shared parameters and client‐specific distributions with personalized parameters. Extensive experiments on eight benchmark datasets demonstrate that SeqFedRPC surpasses eight SOTA methods, with each client personalizing less than 10% of the total parameters on all datasets at α=0.1$$ \alpha =0.1 $$ .
Peng Mao, Tian Du, Zhonghui Wu et al.· Transactions on Emerging Tel...· 0 citations
As Satellite Edge Computing (SEC) emerges as a computing backbone of 6 G IoT, its feasibility is threatened by frequent handovers that disrupt task continuity and cause deadline violations in AI-driven workloads. Conventional communication-centric handover strategies overlook the urgency of computational tasks, further exacerbating these issues. To address this gap, we propose the <bold>C</bold>omputation <bold>A</bold>ware <bold>C</bold>onditional <bold>H</bold>andover <bold>O</bold>ptimization (CACHO) framework, a non-intrusive solution fully compatible with 3GPP Conditional Handover (CHO). The framework integrates model partition driven delayed execution to enable seamless migration via inter-satellite tensor transmission, timeout-risk predictive proactive triggering to protect at-risk tasks from overloaded satellites, and task-load-aware target selection to balance computational workloads and service popularity. These modules are orchestrated within standardized CHO workflows without altering the core logic, ensuring deployability in existing satellite network systems. Simulations based on real Starlink constellation data demonstrate that our approach reduces task timeout rates by 29.95–51.02<inline-formula><tex-math notation="LaTeX">$\%$</tex-math><alternatives><mml:math><mml:mo>%</mml:mo></mml:math><inline-graphic xlink:href="xu-ieq1-3674471.gif"/></alternatives></inline-formula> compared to the 3GPP standard CHO and several state-of-the-art baselines, while maintaining handover success rates and minimizing extra handovers needed. Beyond performance gains, this work provides the first blueprint for embedding task life-cycle awareness into handover workflows, bridging mobility management and computation guarantees, and advancing the feasibility of robust SEC for mission-critical 6G IoT applications.
Chuxing Fang, Changqiao Xu, Zitong Li et al.· IEEE Transactions on Mobile...· 0 citations