Federated learning (FL) enables collaborative model training without moving raw data, but it is limited by heavy and uneven network traffic, particularly when many edge devices share updates. In this work, we present two novel quantization schemes, Frequency Conditional Quantization (FCQ) and Linear-Frequency Projection (LFP), which are frequency-domain-based compression methods tailored for FL systems. At each training round, each client quantizes its model weights using the most suitable scheme according to its current bandwidth and traffic conditions. We validate the proposed adaptive compression framework through extensive experiments on FEMNIST and CIFAR-10 and CIFAR-100 image classification tasks. Across the methods, the proposed adaptive algorithm attains accuracy close to the uncompressed baseline (FEMNIST: 0.938 vs. 0.943; CIFAR-10: 0.811 vs. 0.814) while reducing latency relative to high-accuracy baselines (adaptive: 0.13 s vs. FP8: 0.35 s). A stress test under three bandwidth regimes (Balanced, Mostly Slow, Mostly Fast) indicates stable performance (94.43%, 92.54%, and 93.22% accuracy, respectively) with modest variation in total duration. The results support frequency-domain quantization combined with client-aware scheme selection as a practical approach to reduce communication and per-round time while maintaining model quality in bandwidth-constrained federated networks.
Ahmed et al. (Thu,) studied this question.