Owing to strong capabilities in capturing interactions among objects and concepts, graph data has been treated as an important type of information collected by smart devices in Industrial Internet of Things (IoT), and the distributed training of graph learning models over these devices brings fundamental supports for intelligent services and operations. However, different IoT devices may collect Non-IID graph data due to different roles in the system, and suffer poor performance when only one unified instance of model is trained. Besides, IoT devices usually belong to different communities in Industrial IoT, such that each community pursues both optimized and rational performance when joining in the training process. Considering both challenges, this article proposes a novel meta-computing framework for federated graph learning in Industrial IoT. A collaborative resource allocation task is formulated where devices belonging to different communities adopt limited resources to participate in the training of multiple instances either within or across communities. Two algorithms are introduced for adaptive and rational resource allocation based on whether devices are owned by single or multiple communities. Both algorithms provide guaranteed performance on efficiency and effectiveness, and the fairness among IoT devices are proved. Finally, extensive numerical results have demonstrated the performance of the proposed framework in handling collaborative graph model learning within Industrial IoT.

Distributed mobile devices collect unlabeled graph data from environment. Introducing popular graph contrastive learning (GCL) methods can learn node representations better. However, training high-performance GCL requires large-scale data and graph data collected by the single device is insufficient. Meanwhile, transmitting local data for centralized training suffers from non-negotiate privacy leakage and bandwidth consumption. Federated learning (FL), as a distributed learning paradigm, is commonly used for such issues. Nevertheless, direct combination of FL and GCL struggles to supplement global graph information. This absence results in neighbor information missing, thus causing the local GCL to learn biased node representations. Moreover, the combination also triggers potential gradient explosion owing to the lack of unified learning criteria. In this paper, we propose a federal GCL framework that complements missing structural information and provides unified learning criteria. The key idea is to achieve cross-client node alignment on server through local graph structural importance to reason about the global graph information. We design a hierarchical structural importance scoring method to comprehensively evaluate structural importance, thus server performs effective cross-client aggregation while maintaining local graph privacy. We demonstrate the security and prove the bandwidth-reducing advantage of the proposed framework. Extensive experiments on 3 datasets show the superior performance of our method.

The Internet of Thing systems have contributed a considerable scale of graphs via numerous devices, like the network topology. These graph data are used in wireless communication and mobile computing for a variety of tasks such as traffic prediction and vehicle communication. Therefore, the analysis of these graph data has shown great significance towards the management of IoTs. Compared to other distributed computing scenarios, IoTs require a wider range and more frequent transmission of data. Considering the constrained communication resources and concerns about data privacy, IoT devices tend to keep the graph data locally. Therefore, it is meaningful to adopt the idea of Federal Graph Neural Network for IoTs, where the fancy Graph Neural Network can be implemented in a distributed manner over graphs in IoTs. This paper proposes a sampling-based framework towards the purpose. It assumes the graphs are vertically partitioned over devices, which means all devices hold an identical set of vertices and edges, and each owns a subset of features. It fits the fact that different devices own heterogeneous sensing capabilities. Then a sampling-based method is proposed for multi-round training of Federal Graph Neural Network, and only partial devices join in the training in each round. The sampling strategy applies the local accuracy of validation of each device to adjust the sampling probabilities accordingly, so as to accelerate the convergence of the global model. Finally, extensive analysis and numerical evaluation verify the advancement of the framework.

Pan-sharpening, a task involving information fusion, entails merging panchromatic (PAN) images with high spatial resolution and low-resolution multispectral (LRMS) images in order to obtain high-resolution multispectral (HRMS) images. Due to deep learning’s excellent regression capabilities, it has recently become the dominating technique for this assignment. Meanwhile, the development of the transformer, a novel deep learning architecture for natural language processing, has provided researchers with new insights. In this letter, we seek to extend transformer’s excellent mechanisms to pixel-level fusion challenges. We designed a parallel convolutional neural network structure for learning both the regions of interest from the LRMS images and the residuals required for regression to HRMS images. Then, in our proposed pixelwise attention constraint (PAC) module, the residuals will be changed utilizing the learned region of interest. In addition, we presented a novel multireceptive-field attention block (MRFAB) to frame our network. Experiments on two datasets also show that our work is better than the mainstream algorithms at both indicators and visualization.