Split Learning with MobileViT: A Performance Evaluation of Communication and Inference

In traditional client–server (CS) architectures, transmission latency is often the main bottleneck in object detection. To address this issue, we study partitioned model inference, which uses the spatial downsampling property of convolutional neural networks (CNNs) to reduce the amount of data that must be sent over the network. Concretely, we integrate the lightweight MobileViT backbone into the YOLO framework and apply partitioned inference so that the client can compress the transmission payload with only a small computational cost. We design an optimized partition strategy that reduces the total inference latency by 9.7% while also offloading part of the computation from the server to the client. A combination of theoretical modeling and experiments under real network conditions shows the effectiveness of the proposed system. The results indicate that our method can balance inference efficiency and communication bandwidth, and provides a practical basis for adaptive model deployment in real-world network environments.