Real-time Network-Aware Compression for Efficient Split Learning in Distributed Systems

The rapid growth of distributed machine learning, particularly Split Learning, has intensified the challenge of communication overhead, e.g., in networking environments with limited bandwidth and fluctuating traffic conditions. Existing approaches typically apply a fixed compression level throughout training, failing to account for network traffic variability and the trade-off between communication efficiency and model accuracy.In this paper, we propose a dynamic compression technique for Split Learning that adapts in real-time to network traffic conditions by combining a network-aware drop rate with quantization. The drop rate determines which intermediate features are transmitted based on current bandwidth and latency. On the other hand, quantization reduces the precision of the transmitted data, jointly minimizing communication without significantly impacting model performance.Our approach also provides an implicit regularization effect, improving generalization by preventing the model from memorizing irrelevant patterns. Extensive experiments on a real-world testbed with Non-Independent and Identically Distributed data across heterogeneous devices demonstrate compatibility with multiple Split Learning architectures. They also show that key network performance indicators (Round Trip Time, Transmission Control Protocol retransmissions, and bandwidth utilization) remain stable under heavy congestion.Overall, our method can reduce data transfer by up to 82% during training and inference, while controlling accuracy degradation, making it highly suitable for communication-constrained environments such as 5G or 6G networks.