Danyang Zheng

Federated learning (FL) offers a promising approach for collaboratively training models across distributed clients without centralizing raw data. However, practical deployment remains challenging due to two intertwined bottlenecks: data heterogeneity and random client participation, which jointly make aggregated server updates spectrally concentrated and weaken effective optimization progress. Existing methods typically address heterogeneity through client-side regularization, representation alignment, or aggregation-weight adjustment, but rarely examine the singular-value geometry of the raw update matrix actually applied by the server. To address this gap, we propose Spectral Rebalancing Aggregation (SRA), a lightweight server-side post-aggregation method that mitigates aggregated-update spectral concentration while preserving the standard FL workflow. SRA first identifies that stronger non-IID partitioning decreases spectral flatness and increases condition numbers, indicating that update magnitude concentrates in a few dominant singular directions. Based on this observation, SRA applies blockwise Newton-Schulz spectral rebalancing to strengthen under-updated singular directions and further introduces Frobenius-norm rescaling to prevent update-scale drift. Experiments demonstrate that SRA improves late-stage accuracy by up to 21.44 percentage points over representative baselines while maintaining improved training stability.

Artificial Intelligence (AI) techniques are driving increasing volumes of reliability-sensitive Service Function Chain (SFC) requests from applications such as autonomous driving, smart manufacturing, and intelligent healthcare. These applications require not only feasible deployment but also continuous service availability under failures, disruptions, and uncertainty. Existing reliable SFC deployment studies have largely relied on rule-based approaches, which provide explicit formulations and interpretable decision rules but often struggle with dynamic, large-scale, and multi-objective requirements involving reliability, latency, cost, and security. AI-driven solutions therefore offer a promising direction for enabling adaptive and data-driven deployment decisions. This survey first clarifies the workflow of SFC deployment and the distinction between conventional and reliable SFC deployment. We then review rule-based reliable SFC deployment, analyzing its major design rationales, representative solution paradigms, and limitations. Building on this analysis, we examine AI-driven methodologies for reliable SFC deployment, with a particular focus on deep reinforcement learning approaches, which dominate the existing literature. Finally, we discuss open challenges and future directions toward adaptive, robust, and scalable reliable SFC deployment.

Incomplete multi-view clustering (IMVC) has recently attracted increasing attention and achieved notable progress in computer vision. However, existing IMVC approaches still face several critical challenges. First, most methods fail to capture the inherent nonlinear structures of real-world data. Second, they fail to sufficiently exploit low-rankness and smoothness of multiple views. To overcome these limitations, we propose a novel method, termed Incomplete Multi-View Kernel Subspace Clustering with Tensor-Correlated Total Variation Regularization (KSC-TCTV), which integrates the ability of kernels to capture nonlinear separability with the strength of tensors in characterizing high-order correlations. Specifically, KSC-TCTV first effectively models nonlinear structures by embedding data into a kernel Hilbert space. And then, we introduce a log-based tensor-correlated total variation (Logt-CTV) regularizer in the kernel space, which jointly enforces global low-rankness for inter-view dependency modeling and local smoothness for intra-view structure preservation. The Logt-CTV employs logarithmic non- convex relaxation to mitigate the estimation bias. Experiments on several public benchmark datasets demonstrate that KSC-TCTV outperforms the state-of-the-art IMVC methods.

Mixture-of-Experts (MoE) inference differs from conventional centralized models by activating only a small subset of experts for each token and allowing expert models (EMs) to be distributed across edge servers. This sparse activation paradigm makes MoE attractive for edge inference, but also creates a new deployment challenge because expert placement must jointly optimize computation, memory, bandwidth, and security under heterogeneous expert-level security constraints. In this paper, we study the cost-efficient deployment of MoE inference in edge networks under expert-level security requirements. We formulate the Expert-level Security-aware MoE Deployment (ESMD) problem by jointly modeling inner-service resource sharing (ISRS), server-level security enhancement (SLSE), and edge resource constraints in the overall deployment cost. To solve ESMD, we develop a Composite Deployment Loss (CDL) metric and an Expert Activation Graph (EAG) to capture the coupled effects of security provisioning and resource sharing during expert placement. Based on these designs, we propose CDL-SPE, a cost-oriented deployment algorithm for MoE inference in edge networks.Under practical assumptions, we show that CDL-SPE achieves a logarithmic approximation ratio. Extensive simulations demonstrate that CDL-SPE consistently reduces deployment cost and improves resource-sharing efficiency compared with the state-of-the-art benchmarks.

Deploying ultra-reliable and low-latency communication (URLLC) service function chains (SFCs) is imperative for applications demanding stringent latency and reliability performances. In these applications, ensuring uninterrupted service hinges on establishing fault-disjoint primary and backup service function paths (SFPs). However, existing techniques for deploying SFCs fall short in optimizing latency differentials between the primary and backup SFPs, posing risks of service disruptions in critical URLLC applications like remote surgery, smart factory, and unmanned vehicle systems. In this work, we investigate pioneering techniques to efficiently optimize the primary and backup SFP latencies while minimizing their differentials. We formally formulate the problem of ultra-reliable and low-latency SFC deployment (URLLC-SD) and show its NPhardness. We develop an innovative algorithm, the Yen-based SFP Identification in Layered Graph (YANG), which optimizes the equal-weight composite latency objectives with symmetric QoS/SLA for primary and backup SFPs at the expense of runtime complexity. Through extensive simulations, we demonstrate the YANG's superiority, surpassing state-of-the-art benchmarks. In particular, YANG achieves the highest acceptance rates under specific constraints on SFP latency and latency differentials. Furthermore, our analysis reveals interesting insights into the selection of good path candidates to optimize URLLC-SD.