Original title: Have Your Cake and Eat It Too: Toward Efficient and Accurate Split Federated Learning
Authors: Dengke Yan, Ming Hu, Zeke Xia, Yanxin Yang, Jun Xia, Xiaofei Xie, Mingsong Chen
The article tackles Split Federated Learning’s (SFL) drawbacks in AIoT systems: low accuracy and efficiency due to data differences and slow devices. They introduce Sliding Split Federated Learning (S$^2$FL), a smart solution. S$^2$FL adapts by splitting models dynamically based on each device’s power, tackling slow devices. It also blends data from devices with diverse information, creating better batches for learning. This helps handle data differences. Their tests show S$^2$FL outperforms regular SFL, boosting accuracy by up to 16.5% and speeding up training by 3.54 times. Essentially, S$^2$FL is a smarter way for devices to work together. It tailors the training to each device’s speed and mixes their data to make learning more efficient and accurate, making AIoT systems smarter and faster.
Original article: https://arxiv.org/abs/2311.13163