Lapa: Multi-Label-Flipping Adversarial Attacks on Graph Neural Networks

In recent years, the robustness of graph neural networks (GNNs) adversarial attacks has been favored by researchers, and related research has intensified. However, existing graph adversarial attacks are mainly performed by modifying the topology of the graph and node feature vectors, and attacks from the labels of nodes are rarely considered, especially for multi-label graph adversarial attacks, where an attacker can manipulate an obscure portion of the training labels and flip the sample labels in the training data to other classes, resulting in incorrect predictions of the retrained GNNs model. In this work, we present graph adversarial attack against multi-label flipping. We propose an effective attack model-Lapa, based on an adaptive genetic algorithm that searches the solution space for suitable adversarial samples to obtain near-optimal solutions. The adaptive genetic algorithm performs crossover and mutation operations utilizing adaptive probabilities. We demonstrate the effectiveness of our proposed Lapa attack model on GNNs after extensive experiments on four real datasets.