Análise de desempenho da quantização em soluções de aprendizado dividido federado

Abstract

This paper addresses Split Federated Learning (SFL), a technique that enables artificial neural network models to be trained across multiple entities without directly sharing their data, thereby preserving data privacy. By partitioning the model between the client and the server, SFL also helps to reduce the computational load on the client side. However, despite its advantages, using this solution remains challenging in computationally constrained environments, such as mobile devices and IoT devices, especially on the client side. This work analyzes the performance of quantization, a technique that reduces the numerical precision of neural network parameters and activations to lower computational costs for clients during inference in SFL solutions. To this end, an empirical study was conducted in a cloud environment using the VGG11, ResNet18, and MobileNetV2 models over IID and non-IID variations of the CIFAR-10 dataset. The Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) schemes are compared with the full-precision model. The results demonstrate that the gains and losses from quantization are directly linked to the architecture, the choice of the cut layer, and the data distribution. Quantizing the client model accelerated the client’s forward pass by 4.3 on average. The reduction in memory consumption ranged from 26% to 56%, while the decrease in network consumption remained at 75% for all cases. The accuracy loss is also linked to the architecture: while the VGG11 and ResNet18 models maintained an accuracy loss of less than 2%, PTQ on the MobileNetV2 architecture led to losses of up to 22.7%, underscoring the need for QAT.