BDP: Bipartite Graph Adversarial Defense Algorithm Based on Graph Purification

Existing graph adversarial defense algorithms cannot recover the implicit relationships of a bipartite graph that have been disrupted by adversarial attacks. To address this issue, this paper proposes Bipartite graph adversarial Defense algorithm based on graph Purification (BDP) to improve the robustness of bipartite graph embedding algorithms under adversarial attacks. Firstly, we obtain the poisoned graph disrupted by adversarial attack algorithms; secondly, we traverse all the edges in the poisoned graph, and delete the edges whose Jaccard similarity is less than the threshold; thirdly, we traverse the implicit edges in the graph, and recover the important edges deleted by the adversarial attack algorithms according to the implicit edge reconnection rule, so as to obtain the purified graph; Then, we input the purified graph into the training model to get the node embeddings. Finally, we input the node embeddings into the downstream task to detect their performance. The experimental results show that under the same adversarial attack settings, as the most advanced baselines, BDP can significantly improve the performance of bipartite graph embedding algorithms. At the same time, the experimental results also prove that using Jaccard similarity to measure node similarity can make BDP optimally perform. Supplemental materials including codes and datasets are available at https://github.com/DengBW-1998/BDP.