AquaFed: Ascending Quantized Federated Learning on Heterogeneous Devices

Federated Learning (FL) enables privacy-preserving collaboration across distributed edge devices but faces severe challenges in cross-device settings due to stringent memory and computation constraints. While quantization offers a natural remedy, existing techniques primarily target inference acceleration and fail to improve training efficiency. Consequently, current quantized FL frameworks only mitigate communication overhead while suffering from memory inefficiency during on-device training. Moreover, they exhibit Heterogeneity-induced optimization instability, as devices with varying bit-widths jointly participate in federated learning, introducing inconsistent quantization noise that hinders convergence and degrades final accuracy. In this work, we propose AquaFed, a memory-efficient Ascending quantized Federated learning framework tailored for heterogeneous edge environments. We first introduce Memory-Efficient Quantization (MEQ), which quantizes both weights and activation caches during training to drastically reduce peak memory usage and memory I/O. To address instability arising from heterogeneous precision, we theoretically establish the Quantization Noise Equivalence (QNE) Hypothesis: Quantization error exhibits a behavior analogous to the stochastic noise induced by a large learning rate in quantized federated learning. Guided by QNE, AquaFed adopts an ascending quantization schedule, allowing low-bit devices to lead early coarse-grained learning, while high-bit devices refine the model in later stages for stable convergence. Extensive experiments on vision and language benchmarks demonstrate that AquaFed consistently delivers superior accuracy, generalization, and memory efficiency compared to state-of-the-art methods. On CIFAR-10 with ResNet-18, AquaFed achieves a 20% accuracy improvement and 39% memory saving over the best existing framework. © 2013 IEEE.