WZWenxuan Zhou

Wenxuan Zhou

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Wenxuan Zhou has a total of 9 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    FedCSN: Mitigating Skewed Feature via Class-wise Selective Normalization in Federated Learning

    by Haojie Sun, Zhihao Qu, Yanyan Wang, Wenxuan Zhou, Ninghui Jia, Baoliu Ye

    Federated learning (FL) enables multiple distributed clients to collaboratively train a shared model without directly exchanging raw data. However, heterogeneous feature distributions across clients often lead to inconsistent local optimization and degraded model performance. Existing normalization-based augmentation methods attempt to alleviate this issue through cross-client statistic sharing. Nevertheless, most existing methods do not explicitly consider class consistency during normalization perturbation, where statistics from different classes may be indiscriminately applied to local samples. Through our investigation, we observe that class-consistent statistic sharing can improve model performance, while cross-class statistic perturbation may introduce noisy augmentation under heterogeneous edge environments. Motivated by these observations, we propose FedCSN, a lightweight normalization-based augmentation framework based on class-wise selective normalization for FL under feature distribution skew. Specifically, FedCSN constructs class-wise normalization statistics from local datasets and adopts an adaptive dominant-class statistic selection strategy to improve the reliability of uploaded class-wise statistics. During local training, each sample is augmented using normalization statistics associated with its class label, thereby reducing cross-class statistic mismatch. Extensive experiments demonstrate that FedCSN achieves the highest dataset-level average accuracy among the evaluated methods across all three datasets using both AlexNet and ResNet18 under severe Non-IID settings. In particular, using AlexNet, FedCSN improves the average accuracy on Office-Caltech-10 by 6.12 percentage points compared with the state-of-the-art normalization-based augmentation method. The source code is publicly available at https://github.com/MereIN0211/FedCSN.

  • Feature Map Compression for Split Learning in Mobile Wireless Environment: Theory and Practice

    by Wenxuan Zhou, Zhihao Qu, Baoliu Ye, Shen-Huan Lyu, Miao Cai, Mingtao Ji, Song Guo

    Split Learning (SL) enables collaborative model training by distributing computation between resource constrained mobile edge devices and a central server. In this distributed framework, mobile edge devices compute the early layers of the model and transmit feature maps to the server. However, due to practical mobile wireless bandwidth limitations, these feature maps should be compressed before transmission. In this paper, we propose a lightweight Mask-Encoded Spar sification (MS) compressor that enhances Top-k sparsification with a narrow-bit mask, significantly reducing compression error while improving computational and communication efficiency. Our theoretical analyses reveal that feature map compression generally induces estimator bias in the gradients. Furthermore, we establish the relationships between compression error and convergence, as well as between the choice of cutlayer and output error. The first relationship demonstrates that reducing compression error facilitates convergence, confirming the advantage of MS, whereas the second provides theoretical guidance for determining the cutlayer, indicating that earlier layers are more sensitive to compression errors. Extensive experiments on vision and language models demonstrate that our method outperforms existing compression techniques in terms of compression error and model performance, achieving up to a 21.38× speedup in communication efficiency under wireless environments.