Zasnova in implementacija heterogenega testnega okolja za zvezno učenje

This master's thesis presents the design and implementation of CoLExT-A, a heterogeneous federated learning (FL) testbed utilizing Android smartphones. FL enables decentralized model training on edge devices, enhancing data privacy by keeping data local. CoLExT-A addresses the gap between simulated FL environments and real-world applications by providing a platform for testing FL algorithms under realistic conditions (device variability and heterogeneity). The testbed supports customizable FL strategies and offers comprehensive metric tracking, including hardware metrics like CPU/GPU utilization and power consumption. Experimental results demonstrate the testbed's ability to identify real-world challenges in FL, such as straggler effects due to device heterogeneity. CoLExT-A serves as an open-source resource for the FL research community, facilitating the development and evaluation of robust FL systems.