<p>Due to the heterogeneity and vulnerability of Internet of Things (IoT) devices, their widespread applications pose considerable challenges to network security. The very premise of network security is accurate identification of IoT devices that access the network. However, the existing device identification methods demonstrate poor performance in extracting fine-grained features of devices and often fail to differentiate similar devices. To address this limitation, we propose an IoT device identification method based on the residual-connected Capsule Network (RC-CapsNet). In this method, we introduce the Residual-Resolution Network (Res2Net) block to the network, thus enhancing its multi-scale feature extraction capacity and enabling the capture of device features from a broader perspective. This method also employs the channel and spatial attention mechanism to highlight key features and improve the representation ability of the model. Additionally, the features from different stages are fused to alleviate feature degradation during training. Finally, capsule vectors are employed to further learn inter-feature spatial relationships and deeper information. Experiments demonstrate that our method yields a remarkable improvement in identification accuracy. Specifically, our RC-CapsNet model achieves a precision of 96.90%, a recall of 96.74%, an F1-score of 96.73%, and an accuracy of 96.74%, outperforming all other models for all the 4 indicators. In comparison with CapsNet, RC-CapsNet boosts the precision by 8%, the recall by 7.85%, F1-score by 7.94%, and the accuracy by 7.85%. These numerical results clearly prove its superiority over existing methods and establish it as an outstanding solution for IoT device identification.</p>