Although the efficient application of emerging technologies such as cloud computing, big data, and image recognition in the supervision of new power systems has effectively improved the supervisory effectiveness of power systems, it has also increased the difficulty of network security protection of power systems while improving work efficiency. In this paper, we evaluate the robustness of a graph convolutional neural network for new power systems to address the problem that artificial intelligence models for power image classification and recognition are susceptible to adversarial perturbation deception. Firstly, this paper achieves high-confidence spoofing of targets with two hops and more than two hops based on node classifiers. Firstly, this paper achieves high-confidence spoofing of targets with two hops and more than two hops based on node classifiers. Secondly, this paper establishes the POISONPROBE algorithm to improve the robustness of the attack model by searching for smaller perturbations within a longer distance from the target. Finally, experiments based on two datasets are conducted in this paper to demonstrate that the proposed POISONPROBE algorithm can be used to improve the robustness of the attack model by searching for smaller perturbations. proposed POISONPROBE algorithm has a high attack success rate. The proposed attack method in this paper can be used in future grid defense experiments to study graph convolutional neural networks with antagonistic robustness.