Dynamic Single-Step Synthetic Features Compression for Communication-Efficient Federated Learning

Federated Learning (FL) protects data privacy through distributed training, yet its high communication overhead restrictions applicability in resource-constrained environments. This paper proposes a communication-efficient FL method— Dynamic Single-Step Synthetic Features Compressor (D-3SFC)— which generates compact synthetic datasets via single-step optimization. By dynamically adjusting the number of synthetic samples based on gradient magnitude, client heterogeneity, and training loss, D-3SFC effectively maintains model performance while significantly reducing communication costs. To further improve convergence stability under non-IID conditions, D-3SFC incorporates heterogeneity-aware strategies, including weighted gradient aggregation and momentum-based updates. Experiments are conducted using MLP and MNISTNet on the MNIST and FMNIST datasets, with a Dirichlet parameter α = 1.0 to simulate high heterogeneity. Experimental results demonstrate that D-3SFC reduces communication overhead by 85.6 times compared to FedAvg, achieving server accuracies of 90.03% and 85.63%, respectively—outperforming baseline methods such as SignSGD, Top-K, and FedSynth, with a cosine similarity of 0.89. The single-step optimization design substantially lowers computational complexity, making D-3SFC well-suited for low-resource devices. This work offers a novel and practical solution for communication-efficient federated learning in constrained environments.