Abstract | ||
---|---|---|
Graph neural networks can learn graph structure data directly and mine its information, which can be used in drug research and development, financial fraud prevention, and other fields. The existing research shows that the graph neural network is lacking robustness and is vulnerable to attack by adversarial examples. At present, there are two problems in the generation of confrontation examples for graph neural networks. One is that the properties of graph structure are not fully used to describe the antagonistic examples, the other is that the gradient calculation is linked with the loss function and not directly linked with the properties of graph structure, which leads to excessive search space. To solve these two problems, this paper proposes a graph structure data confrontation example generation scheme based on graph theory measurement. In this paper, the average distance and clustering coefficient is used as the basis for each step of disturbance, and the counterexamples are generated under the premise of keeping the data characteristics. Experimental results on small-world networks and random graphs show that, compared with the previous methods, the proposed method makes full use of the nature of graph structure, does not need complex derivation, and takes less time to generate confrontation examples, which can meet the needs of iterative development. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1007/978-3-030-97774-0_36 | Smart Computing and Communication |
Keywords | DocType | Volume |
Adversarial examples, Graph convolutional network, Graph theory metrics | Conference | 13202 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Wenyong He | 1 | 0 | 0.34 |
Mingming Lu | 2 | 343 | 30.42 |
Yiji Zheng | 3 | 0 | 0.34 |
Naixue Xiong | 4 | 2413 | 194.61 |