Improving the adversarial transferability with relational graphs ensemble adversarial attack

Abstract

In transferable black-box attacks, adversarial samples remain adversarial across multiple models and are more likely to attack unknown models. From this view, acquiring and exploiting multiple models is the key to improving transferability. For exploiting multiple models, existing approaches concentrate on differences among models but ignore the underlying complex dependencies. This exacerbates the issue of unbalanced and inadequate attacks on multiple models. To this problem, this paper proposes a novel approach, called Relational Graph Ensemble Attack (RGEA), to exploit the dependencies among multiple models. Specifically, we redefine the multi-model ensemble attack as a multi-objective optimization and create a sub-optimization problem to compute the optimal attack direction, but there are serious time-consuming problems. For this time-consuming problem, we define the vector representation of the model, extract the dependency matrix, and then equivalently simplify the sub-optimization problem by utilizing the dependency matrix. Finaly, we theoretically extend to investigate the connection between RGEA and the traditional multiple gradient descent algorithm (MGDA). Notably, combining RGEA further enhances the transferability of existing gradient-based attacks. The experiments using ten normal training models and ten defensive models on the labeled face in the wild (LFW) dataset demonstrate that RGEA improves the success rate of white-box attacks and further boosts the transferability of black-box attacks.