Key points are not available for this paper at this time.
Federated learning (FL) is envisioned as a pioneering framework for the distributed training of artificial intelligence models in mobile edge computing (MEC) environments. Traditional MEC-empowered FL approaches commonly neglect the inherent competition for computation resources between uncertain user-own tasks and FL training activities on individual Internet-of-Things (IoT) devices. Meanwhile, these approaches fail to address the personalized reward perception that dominates the active participation of IoT devices in FL training. As a result, both the resource utilization on IoT devices and the overall performance of FL models are significantly degraded in practical MEC systems. To address these challenges, this paper investigates the personalized FL with state-adaptive IoT device scheduling in MEC scenarios. We first develop a collaborative device-state estimation method to effectively capture the uncertainty in future states of FL candidates. Subsequently, we design a user-personality inspired degree-of-satisfaction (DoS) prediction scheme to quantify the impact of computation resource competition on the satisfaction levels of personalized FL participants. Building on these efforts, we propose a state-adaptive IoT device scheduling technique to optimize the accuracy of FL models at the offline stage. An FL runtime management policy is also designed to deal with the timing failures of unsuccessful return of local training results at the online stage. Evaluations show that our approach enhances the accuracy of WideResNet FL model by up to 35.96% on CIFAR-10 and EuroSAT datasets. Our source code is available at https://github.com/superguymj/ACE.
Mai et al. (Fri,) studied this question.