Performance Analysis of Quantization in Federated Split Learning

• Claro Henrique S. Sales UFC
• Francisco Bruno Heron de Carvalho Junior UFC
• Allberson de Oliveira Dantas UNILAB

Resumo

Collaborative learning techniques address privacy concerns in deep learning when working with large datasets distributed across multiple entities or computing devices. Among these techniques, Federated Split Learning (FSL) brings Split Learning (SL) techniques to Federated Learning (SL) to minimize computation demands on the client side without compromising privacy requirements. However, even with FSL, client-side inference may remain computationally expensive, which motivates us to evaluate the impact of quantization on FSL in terms of model accuracy and resource usage on the client side.