Evaluation and Optimization of Distributed Machine Learning Techniques for Internet of Things

Federated learning (FL) and split (SL) are state-of-the-art distributed machine techniques to enable training without accessing raw data on clients or end devices. However, their comparative performance under real-world resource-restricted Internet of Things (IoT) device settings remains barely studied. This work provides empirical comparisons FL SL in IoT regarding (i) with heterogeneous distributions (ii) on-device execution overhead. Our analyses this demonstrate that the is better than an imbalanced distribution but worse extreme non-IID distribution. Recently, combined form splitfed (SFL) leverage each benefits (e.g., parallel lightweight computation requirement SL). considers FL, SL, SFL, mounts them Raspberry Pi devices evaluate performance, including time, communication overhead, power consumption, memory usage Besides evaluations, we apply two optimizations. First, generalize SFL by carefully examining possibility a hybrid type model at server-side. The generalized merges sequential (dependent) (independent) processes thus beneficial system large scale devices, specifically server-side operations. Second, propose pragmatic substantially reduce overhead up four times for (generalized) SFL.