Research on Graph Injection Attack Method Based on Multi Objective Feature Optimization in the Electric Power Field

In the field of electricity, in response to the problem of graph neural networks having both significant performance and poor robustness, this may lead to poor recommendation performance or affect social evaluation, posing a threat to the integrity and security of social networks. Therefore, this article proposes a graph injection attack method based on multi-objective feature optimization, which selects multiple target nodes with similar features through clustering target selection strategy and directly connects the injection nodes to the target nodes, ensuring the effective propagation of disturbance information. This operation can enable the disturbance information attached to the injection node to propagate more directly to each target node. For the generation of perturbation features, the selected target node is used to aggregate its own information and neighboring node information to generate it. In order to enhance the perturbation ability of perturbation features, a zero order random coordinate descent algorithm is applied to optimize the generated features for discrete feature maps.