Mind Your Solver! On Adversarial Attack and Defense for Combinatorial Optimization

It is worth noting that many challenging task not only in its inherent CO problems can be essentially formulated as a graph problem complexity (e.g. NP-hard) but also the possible (Khalil et al., 2017; Bengio et al., 2020), hence it is sensitivity to input conditions. In this paper, we attractive and natural to modify the problem instance by take an initiative on developing the mechanisms modifying the graph structure, to generate more test cases for adversarial attack and defense towards combinatorial for solvers. In fact, vulnerability can often be an inherent optimization solvers, whereby the solver challenge for CO solvers since the problem is often strong is treated as a black-box function and the original nonlinear and NP-hard. From this perspective, we consider problem's underlying graph structure (which is attack and defense CO solvers in the following aspects.