Certified Robustness on Structural Graph Matching

Huaqing Shao, Lanjun Wang, Yongwei Wang, Qibing Ren, Junchi Yan

Keywords: Structural graph matching (GM), certified robustness, randomized smoothing, joint Gaussian distribution

TL;DR: We are the first to define certified robustness on GM and design a new certification strategy using a joint smoothing distribution to maximize certified region.

Abstract: The vulnerability of graph matching (GM) to adversarial attacks has received increasing attention from emerging empirical studies, while the certified robustness of GM has not been explored. Motivated by randomized smoothing, we are the first to define certified robustness on GM and design a new certification strategy called Structure-based Certified Robustness of Graph Matching (SCR-GM). Structural prior information of nodes is used to construct a joint smoothing distribution matrix with physical significance, which certifies a wider range than those obtained by previous iterative optimization methods. Furthermore, we propose a certified space that can be used to derive a strictly certified radius and two radii for evaluation. Experimental results on graph matching datasets reveal that our strategy achieves state-of-the-art $\ell_{2}$ certified accuracy and regions.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)