Model Splitting Enhanced Communication-Efficient Federated Learning for CSI Feedback

Haijun Zhang, Yanjie Dong, Haijun Zhang 0001, Gaojie Chen, Xiaoyi Fan, Victor C. M. Leung, Xiping HuIEEE Trans. Vehicular Technology 2025 · DOI

TL;DR

This paper introduces CSILocal, a novel federated learning approach for efficient CSI feedback in massive MIMO systems. By splitting the CSI compression model into encoder/decoder head (UE-side) and decoder tail (base station-side) and employing pipeline parallelism, CSILocal significantly reduces communication overhead and training time compared to existing methods, while preserving user data privacy.

What Problem Does This Paper Address?

Modern mobile communication systems, particularly 5G and beyond (mMIMO), rely on accurate channel state information (CSI) at the base transceiver station (BTS) for optimal performance. However, as the number of antennas in mMIMO systems grows, the amount of CSI that user equipments (UEs) need to send back to the BTS becomes a significant communication bottleneck. Existing methods for compressing CSI, such as codebook-based, compressive sensing, and deep learning approaches, often still impose a high communication burden due to frequent parameter exchanges. Federated learning offers a distributed training paradigm, but traditional implementations require UEs to upload large model parameters or gradients, exacerbating the communication overhead and potentially compromising data privacy. This paper addresses the need for a communication-efficient and privacy-preserving federated learning framework for CSI feedback in mMIMO systems.

Methods

The core innovation of this paper is the CSILocal algorithm, which employs a model splitting strategy combined with pipeline parallelism.

  • Model Splitting: The CSI compression deep neural network is divided into three parts:
  • Encoder and Decoder Head: These modules reside on each UE, processing their private CSI data.
  • Decoder Tail: This module is located at the BTS. This split allows UEs to only exchange "local smashed data" (intermediate activations) with the BTS, which is significantly smaller than full model parameters, thus reducing communication overhead and preserving user data privacy.

  • Pipeline Parallelism: To further accelerate training, the decoder tail at the BTS is split into multiple stages and processed in a pipeline fashion across multiple GPUs. This allows the BTS to process multiple micro-batches of smashed data concurrently during both forward and backward passes, reducing the overall training duration.

The training objective is to minimize the normalized mean squared error between the reconstructed CSI and the original CSI. The algorithm leverages the Adam optimizer for parameter updates.

Datasets

The paper uses a dataset derived from the CsiNet [10] benchmark. The CSI matrices are converted to the angular-delay domain via two-dimensional discrete Fourier transform. The experiments utilize a training set of 100,000 samples, a validation set of 30,000 samples, and a testing set of 20,000 samples. Each data sample has a size of 32x32.

A synthetic dataset was also constructed to evaluate the impact of data heterogeneity (Non-IID setup). This dataset consists of 130,000 training samples and 20,000 testing samples, with an indoor-to-outdoor CSI data ratio of 1:1. The CSI data distribution is varied across 10 UEs to simulate non-IID conditions.

Key Insights and Ideas

  • Privacy-Preserving Federated Learning: By splitting the model and only exchanging intermediate activations, CSILocal effectively preserves the privacy of individual UE's CSI data, a critical concern in federated learning.
  • Communication Efficiency is Paramount: The paper strongly emphasizes that for large-scale mMIMO systems, reducing the sheer volume of exchanged data is more critical than minor improvements in reconstruction accuracy if those improvements come at a high communication cost.
  • Pipeline Parallelism for Server-Side Speedup: While UEs benefit from reduced communication, the BTS can become a bottleneck. Pipeline parallelism on the BTS's GPUs is a clever way to accelerate the processing of aggregated data, directly impacting training time.
  • Trade-off Management: CSILocal demonstrates a good balance between CSI reconstruction accuracy and communication efficiency, particularly under moderate compression.
  • Data Heterogeneity is a Major Challenge: The Non-IID scenario consistently shows higher testing loss, highlighting the difficulty of training robust models when data distributions vary significantly across UEs. However, increasing encoding dimensions can help mitigate this.

Key Results

  • Reduced Communication Overhead: CSILocal significantly reduces the number of exchanged parameters compared to benchmarks like FedAvg and FedProx. For instance, to achieve a testing loss of 1.3 × 10⁻³, CSILocal requires only 19.66 billion exchanged parameters, while FedAvg and FedProx require 47.06 billion and 38.24 billion, respectively.
  • Reduced Training Time: Pipeline parallelism leads to noticeable reductions in training time per iteration. For example, with a mini-batch size of 800 and an encoding dimension of 256, wallclock durations are reduced by approximately 21.85% for indoor data and 21.98% for outdoor data compared to a non-pipeline approach. These speedups are expected to increase with more UEs and larger batch sizes.
  • Competitive Reconstruction Accuracy: CSILocal achieves competitive CSI reconstruction accuracy, often matching or slightly outperforming other methods, especially when considering the communication cost.
  • Impact of Compression Ratio: Increasing compression ratios generally lead to improved reconstruction accuracy, but this comes at the cost of information loss. The Non-IID setting is more sensitive to compression.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the Non-IID scenario presents the most challenging conditions, leading to higher testing loss and variability.
  • While CSILocal performs well, the outdoor scenario exhibits higher loss values and greater variability due to channel fluctuations.
  • The performance gains from pipeline parallelism are dependent on the number of available GPUs at the BTS.
  • Future Work:
    • The authors suggest that increasing encoding dimensions can compensate for some of the degradation observed in Non-IID scenarios.
  • The paper implicitly points towards further research into more robust methods for handling extreme data heterogeneity in federated learning for CSI feedback.
  • Exploring more advanced pipeline parallelism strategies or distributed training techniques could further optimize training efficiency.

Summary updated

BibTeX
@article{zhang2025model,
  author = {Haijun Zhang and Yanjie Dong and Haijun Zhang and Gaojie Chen and Xiaoyi Fan and Victor C. M. Leung and Xiping Hu},
  title = {{Model Splitting Enhanced Communication-Efficient Federated Learning for CSI Feedback}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2025},
  doi = {10.1109/tvt.2025.3587821},
}
Figure