Answers are grounded in this paper and cite their sources.
In this work, we consider one challenging training time attack by modifying data with bounded perturbation, hoping to manipulate the behavior(both targeted or non-targeted) of any corresponding trained classifier during time when facing clean samples. To achieve this, we proposed to use an-encoder-like network to generate the pertubation on the training data with one differentiable system acting as the imaginary victim. The perturbation generator will learn to update its weights by the training procedure of the imaginary classifier in order to produce most harmful and imperceivable noise which in turn will lead the lowest power for the victim classifier. This can be formulated into a-linear equality constrained optimization problem. Unlike GANs, solving such is computationally challenging, we then proposed a simple yet effective to decouple the alternating updates for the two networks for. The method proposed in this paper can be easily extended to the specific setting where the attacker can manipulate the predictions of the classifiers according to some predefined rules rather than only making predictions. Experiments on various datasets including CIFAR-10 and a version of ImageNet confirmed the effectiveness of the proposed method empirical results showed that, such bounded perturbation have good regardless of which classifier the victim is actually using on data.
No takes yet. Share an insight, caveat, or question.
Ji et al. (2019) studied this question.