1: \begin{abstract}

2: The graph neural network (GNN) models have presented impressive achievements

3: in numerous machine learning tasks. However, many existing GNN models

4: are shown to be vulnerable to adversarial attacks, which creates a

5: stringent need to build robust GNN architectures. In this work, we

6: propose a novel interpretable message passing scheme with adaptive

7: structure (ASMP) to defend against adversarial attacks on graph structure.

8: Layers in ASMP are derived based on optimization steps that minimize

9: an objective function that learns the node feature and the graph structure

10: simultaneously. ASMP is adaptive in the sense that the message passing

11: process in different layers is able to be carried out over dynamically

12: adjusted graphs. Such property allows more fine-grained handling of

13: the noisy (or perturbed) graph structure and hence improves the robustness.

14: Convergence properties of the ASMP scheme are theoretically established.

15: Integrating ASMP with neural networks can lead to a new family of

16: GNN models with adaptive structure (ASGNN). Extensive experiments

17: on semi-supervised node classification tasks demonstrate that the

18: proposed ASGNN outperforms the state-of-the-art GNN architectures

19: in terms of classification performance under various adversarial attacks. 

20: \end{abstract}