Federated learning (FL) is an emerging distributed and privacy-preserving machine learning framework. However, the performance of traditional FL methods is seriously impaired by the real-world data, which appear to be non-IID. The recent clustered federated learning (CFL) methods eliminate the impact of non-IID data by grouping clients with similar data distribution into the same cluster. Unfortunately, existing CFL methods heavily rely on the pre-setting of the cluster number, failing to achieve adaptive client clustering. We also experimentally observe that imbalanced data largely degrade their correctness of client clustering. In this paper, we present a novel CFL method without manual intervention, named AutoCFL, which can eliminate both effects of non-IID and imbalanced data simultaneously. To deal with imbalanced data, the local training adjustment strategy adaptively adjusts the number of local training epochs for each client. To further improve the clustering correctness and adaptability, the weighted voting-based client clustering strategy automatically groups each client into an appropriate cluster. Extensive experiments are conducted to evaluate the design of AutoCFL with three popular datasets under various data settings. Experimental results demonstrate that AutoCFL outperforms state-of-the-art methods, e.g., on average improving model accuracy by 9.24%, while reducing communication costs by 4.67 in an adaptive manner.

Crowdsensing tasks are usually described by certain features or attributes, and the task assignment essentially performs a matching w.r.t. worker or user’s preference on these features. However, existing matching strategy could lead to an misaligned task coverage problem, i.e., some popular tasks tend to enter workers’ candidate task lists, while some less popular tasks could be always unsuccessfully assigned. To ensure task coverage after the assignment, the system may have to increase their biding costs to reassign such tasks, which causes a high operational cost of the crowdsensing system. To address this problem, we propose to migrate certain qualified workers to the less popular tasks for increasing the task coverage and meanwhile optimize other performance factors. By doing this, other performance factors, such as the task acceptance and quality, can be comparably achieved as recent designs, while the system cost can be largely reduced. Following this idea, this paper presents cTaskMat, which learns and exploits workers’ task preferences to achieve coverage-ensured task assignments. We implement the cTaskMat design and evaluate its performance using both real-world experiments and dataset driven evaluations, also with the comparison with the state-of-the-art designs.

Clustered federated learning (CFL) has proved to be an effective way to alleviate the non-IID (not independently and identically distributed) data challenge, which severely restricts the wider application of federated learning. However, existing approaches either lack adaptability, i.e. , they require an additional number of clusters as a guide when clustering, or lack effectiveness in terms of communication. In this paper, we explore the differences in the ability of different layers in a model to represent non-IID data, and propose a hierarchical CFL approach, named HiCFL , which considers both adaptivity and communication efficiency. The improvement of communication efficiency is due to our proposed novel concept of model stability, which characterizes the variation of model weights during training. Based on model stability, HiCFL can find the proper time to bi-partition the clusters of mobile devices in a hierarchical manner more quickly. We conduct extensive experiments based on popular datasets with various non-IID data settings. The results show that HiCFL achieves excellent performance effectiveness and efficiency. Compared to state-of-the-art approaches, HiCFL can improve the model accuracy by $2.0\% \sim 9.0\%$ , while reducing the communication overheads by $27.3\% \sim 80.6\%$ .

Gesture detection based on radio frequency signals has gained increasing popularity in recent years due to several benefits it has brought, such as eliminating the need to carry additional devices and providing better privacy. In traditional methods, significant breakthroughs have been made to improve recognition accuracy and scene robustness, but the limited computing power of edge devices (the first-level equipment to receive signals) and the requirement of fast response for detection have not been adequately addressed. In this article, we propose a lightweight Wi-Fi gesture recognition system, referred to as WiFine, which is designed and implemented for deployment on low-end edge devices without the use of any additional high-performance services in the process. Toward these goals, we first design algorithms for phase difference selection and amplitude enhancement, respectively, to tackle the problem of data drift caused by user change. Then, we design a cross-dimension fusion method to extract features of finer granularity from information of different dimensions, thus solving the precision problem of feature granularity. Finally, we design a lightweight neural network architecture by leveraging redundancy to reduce computational cost while ensuring satisfactory recognition accuracy. Extensive experimental results show that the proposed system achieves fast recognition of various actions with an accuracy up to 96.03% in 0.19 seconds.

Power fault monitoring based on acoustic waves has gained a great deal of attention in industry. Existing methods for fault diagnosis typically collect sound signals on site and transmit them to a back-end server for analysis, which may fail to provide a real-time response due to transmission packet loss and latency. However, the limited computing power of edge devices and the existing methods for feature extraction pose a significant challenge to performing diagnosis on the edge. In this paper, we propose a fast Lightweight Fault Diagnosis method for power transformers, referred to as LightFD, which integrates several technical components. Firstly, before feature extraction, we design an asymmetric Hamming-cosine window function to reduce signal spectrum leakage and ensure data integrity. Secondly, we design a multidimensional spatio-temporal feature extraction method to extract acoustic features. Finally, we design a parallel dual-layer, dual-channel lightweight neural network to realize the classification of different fault types on edge devices with limited computing power. Extensive simulation and experimental results show that the diagnostic precision and recall of LightFD reach 94.64% and 95.33%, which represent an improvement of 4% and 1.6% over the traditional SVM method, respectively.