Efficient Federated Split Learning on Android Smartphones via Adaptive Offloading Point Mechanism
Pham Duy Thanh, Koji Zettsu · 2025
This paper investigates the development of a practical platform for on-device federated split learning using Android smartphones. We employ an Adaptive Offloading Point (AOP) mechanism that dynamically optimizes the partitioning of machine learning models between smartphone clients and a central server, based on computational capacity and network bandwidth. To this end, we integrate a reinforcement learning approach with Gaussian Mixture Model (GMM) clustering to derive optimal offloading policies that minimize system-wide training time. Our experimental evaluation, conducted on real Google Pixel 8 devices, demonstrates that leveraging smartphones with stronger hardware configurations significantly improves training efficiency compared to conventional edge devices such as the Toradex Apalis modules. These results highlight the importance of mobile device performance in realizing scalable, low-latency federated learning systems at the network edge.