Improving Adversarial Attacks On Graph Neural Networks With Honey Badger Optimization And Efficient Parallel Defense Strategies

It is becoming very useful to learn from graphstructured data using Graph Neural Networks (GNNs); however, these networks are vulnerable to adversarial attacks. There are numerous other distinct ways in which these assaults can be performed, such as evasion, poisoning or perturbation. When such a situation happens the performance of the GNN models drastically reduces with high possibility of wrong forecasts or classifications which may not be required. Regulus and TensorMap are quite insecure; GNNs are more secure in areas of high risk such as healthcare, cybersecurity and financial institutions, thus, it is mandatory to secure GNNs. The goal of this work is to present a wide range of possible adversarial perturbations to provide a broader evaluation of the vulnerabilities of GNNs. In order to make larger-scale GNNs more immune to adversarial effects, the authors present a parallel defensive framework that includes adversarial training, graph sanitization, and resilient aggregation strategies. By incorporating this framework, the model resistance brought about by different adversarial approaches is significantly enhanced, and this is achieved alongside a preservation of adequate performance on classic tasks.