FEDCODE: A Framework and Experimental Study of Model-Update Compression for Communication-Efficient Federated Learning

Abstract

Federated learning trains shared models without centralizing local data, but repeated model exchange can become a bottleneck in bandwidth-, energy-, and latency-constrained edge systems. This paper presents FEDCODE, a communication-aware federated learning simulation framework for reproducible evaluation of update representations and compression methods with explicit accounting of communication volume and runtime under controlled training configurations. Experiments on MNIST, CIFAR-10, and LEAF-based federated datasets show that update representation, quantization scale, and coder choice strongly affect payload size, while lightweight integer and entropy coding can substantially reduce communication. Comparisons under a matched protocol with established FL compression methods show favorable communication–quality trade-offs for the evaluated FEDCODE delta-coding configurations. Adaptive coder-parameter tuning provides automatic parameter selection but does not consistently outperform strong fixed configurations. We further propose Coder-Aware Adaptive Delta Scaling (CAADS), which combines update statistics, drift in update magnitude, and coder-specific code-length estimation to adapt the quantization scale during training. CAADS adjusts the communication–quality trade-off: conservative settings preserve model performance, whereas aggressive scaling further reduces payload at the cost of larger quality degradation. Additional experiments on edge devices show that end-to-end compression benefits depend on device capability and network bandwidth. Overall, effective federated learning compression requires joint consideration of payload, model quality, codec cost, and network conditions.