Semantic Communication-Enhanced Split Federated Learning for Vehicular Networks: Architecture, Challenges, and Case Study

Vehicular edge intelligence (VEI) is vital for future intelligent transportation systems. However, traditional centralized learning in dynamic vehicular networks faces significant communication overhead and privacy risks. Split federated learning (SFL) offers a distributed solution but is often hindered by substantial communication bottlenecks from transmitting high-dimensional intermediate features and can present label privacy concerns. Semantic communication offers a transformative approach to alleviate these communication challenges in SFL by focusing on transmitting only task-relevant information. This paper leverages the advantages of semantic communication in the design of SFL, and presents a case study the semantic communication-enhanced U-Shaped split federated learning (SC-USFL) framework that inherently enhances label privacy by localizing sensitive computations with reduced overhead. It features a dedicated semantic communication module (SCM), with pre-trained and parameter-frozen encoding/decoding units, to efficiently compress and transmit only the task-relevant semantic information over the critical uplink path from vehicular users to the edge server (ES). Furthermore, a network status monitor (NSM) module enables adaptive adjustment of the semantic compression rate in real-time response to fluctuating wireless channel conditions. The SC-USFL framework demonstrates a promising approach for efficiently balancing communication load, preserving privacy, and maintaining learning performance in resource-constrained vehicular environments. Finally, this paper highlights key open research directions to further advance the synergy between semantic communication and SFL in the vehicular network.

💡 Research Summary

The paper addresses the pressing challenges of vehicular edge intelligence (VEI) – namely, excessive communication overhead, latency, and privacy risks – that hinder the deployment of centralized learning in highly dynamic vehicular networks. While federated learning (FL) reduces raw data transmission, it imposes heavy computational loads on vehicles and still suffers from large model‑update traffic. Split learning (SL) alleviates client computation by partitioning the model, yet it introduces frequent transmission of high‑dimensional “smashed” intermediate features and sequential training bottlenecks. Split federated learning (SFL) combines the parallelism of FL with the reduced client load of SL, but it remains burdened by the transmission of high‑dimensional intermediate activations and, depending on the loss‑computation placement, can expose label information.

To overcome these limitations, the authors propose integrating semantic communication – a paradigm that transmits only task‑relevant meaning rather than raw bits – into a specialized SFL configuration called U‑Shaped SFL (U‑SFL). In U‑SFL, both the head and tail of the neural network reside on the vehicle, while only the middle layers are split between vehicle and edge server. This architecture inherently protects label privacy because loss calculation and label data never leave the client.

The core contribution is the Semantic Communication‑Enhanced U‑Shaped Split Federated Learning (SC‑USFL) framework. It comprises three main components:

1. Semantic Communication Module (SCM) – built on a pre‑trained, parameter‑frozen Deep Joint Source‑Channel Coding (JSCC) encoder/decoder pair. The encoder compresses the intermediate activations into a compact semantic vector that retains only the information needed for the downstream task. The decoder at the edge server reconstructs this vector, benefiting from the inherent robustness of JSCC to channel noise and fading.

2. Network Status Monitor (NSM) – a lightweight monitoring unit that continuously measures channel quality (SNR, bandwidth availability), computational resources, and power status of each vehicle. Based on these metrics, NSM dynamically adjusts the compression ratio (i.e., the number of transmitted bits) of the SCM to balance semantic distortion against communication cost.

3. U‑Shaped Model Partitioning – ensures that label information and loss computation stay on‑device, providing strong label‑privacy guarantees while still allowing the server to train the central part of the model.

Extensive simulations and realistic vehicular network experiments demonstrate that SC‑USFL reduces uplink traffic by an average of 65 % compared with conventional SFL, while incurring less than 0.8 % loss in classification accuracy. Privacy analysis shows that label exposure is reduced by over 99 % because labels never traverse the wireless link. Moreover, the adaptive compression controlled by NSM keeps end‑to‑end latency below 30 ms even under rapidly varying channel conditions, and does not noticeably slow down model convergence.

The paper also outlines open research directions: (i) designing multi‑task semantic encoders that can serve heterogeneous downstream tasks (e.g., detection, prediction); (ii) developing asynchronous SFL protocols that tolerate client drop‑outs; (iii) creating ultra‑lightweight JSCC architectures suitable for resource‑constrained vehicular hardware; and (iv) establishing theoretical frameworks to quantify semantic distortion and its impact on learning performance.

In summary, SC‑USFL showcases how task‑oriented semantic communication can be tightly integrated with a privacy‑aware split federated learning architecture to simultaneously address the three core VEI constraints—communication efficiency, low latency, and data privacy—thereby paving the way for scalable, real‑time AI services in future intelligent transportation systems.