Deep graph convolutional networks can mine the deeper information of non-structured data, e.g., capturing complex interactions within sensor topology. However, the over-smoothing problem severely limits the depth of the graph convolutional network (GCN). The initial residual can ensure the nodes retain some initial information during the propagation process, which largely alleviates the over-smoothing problem in deep graph convolutional networks. However, current works only use the grid search method to determine a fixed initial residual ratio, which can not assign the most appropriate initial information for the nodes with different over-smoothnesses. This paper proposes a novel method named Node-Smoothness Based Adaptive Initial Residual Deep Graph Convolutional Network (NSAIR-GCN). Specifically, it can be divided into two processes: (1) considering the over-smoothing from another perspective, i.e., determining whether nodes are over-smoothed based on the difference in node representations before and after updating and identifying those severely over-smoothed nodes; (2) assigning appropriate initial residual ratios to these nodes based on their smoothness. Extensive semi-supervised node classification experiments on several standard datasets have shown that the adaptive initial residual ratio determined by node smoothness performs better than the previous fixed initial residual ratio and achieves the state-of-the-art.

The application of federated learning (FL) in the Internet of Things (IoT) is experiencing rapid growth. FL becomes vulnerable to data poisoning attacks in the IoT environment, characterized by distributed devices, limited resources, and open networks. Since IoT systems rely on measurement data to make intelligent decisions, attackers can deceive the system by flipping the labels of local data, causing it to misclassify source class samples as target labels at a low cost. Existing defense mechanisms primarily rely on similarity analysis to detect malicious behaviors. However, these methods often lack the sensitivity to handle high-dimensional or complex network models, struggling to identify subtle but significant changes in model parameters or updates. Considering this issue, we propose a label-flipping-robust approach to distinguish malicious clients from benign ones. In brief, we adopt Kullback-Leibler (KL) divergence to quantify the difference between client models, which offers an improved ability to capture the subtle variations within the model. Furthermore, we introduce privacy-protecting FL (GAN in FL) to protect users’ privacy while utilizing the shared knowledge in the global generator to support detecting malicious clients. Experimental results on real-world data sets show that our proposed algorithm can effectively defend against label-flipping attacks (LFAs) under various data distributions.