Adversarial Attack Framework on Graph Embedding Models with Limited Knowledge

With the success of graph embedding model in both academic and industry areas, robustness against adversarial attack inevitably becomes a crucial problem learning. Existing works usually perform white-box fashion: they need to access predictions/labels construct their loss. However, inaccessibility makes impractical for real learning system. This paper promotes current frameworks more general flexible sense -- we consider ability various types models remain resilient black-box driven attacks. We investigate theoretical connection between signal processing models, formulate as process with corresponding filter. Therefore, design generalized framework: GF-Attack. Without accessing any labels predictions, GF-Attack can directly on filter fashion. further prove that an effective without assumption number layers/window-size models. To validate generalization GF-Attack, five popular Extensive experiments effectiveness several benchmark datasets.