Graph neural network about member inference attacks

[Chrisable]

GCN is transductive learning, how to judge whether a node on a graph is a member? This paper provides a method of judgment in inductive learning，But,transductive learning?

[ArnabKumarRoy02]

Yes, transductive learning concentrates on generating predictions for the nodes that are already present in the graph, as opposed to inductive learning, where the model must generalise to unknown nodes.

For judging whether a node is a member of the graph, you can:

• Use a trained model to predict the labels of that particular node.
• Set a threshold for prediction (just like confidence of that node being a member), pred <= threshold means not a member.
• Divide the graph into several folds, train on a portion of it, and then evaluate your performance using the remaining data.
• Modifying hyper-parameters like learning rate and layer number, the model's performance on the transductive learning task can be optimised.

These are not different methods, these are the steps to follow to solve you question.

[Chrisable]

It's okay to use this paper directly

[Chrisable]

Looking forward to your reply!

[ArnabKumarRoy02]

Yes, it's okay to use this paper directly. As the paper mentions that GNN models could leak information about the member nodes that were trained on.
So, if your primary question is to check whether a node on a graph is a member or not? You CAN use this paper.