Yet Another Intermediate-Level Attack

Qizhang Li, Yiwen Guo, Hao Chen

Abstract: The transferability of adversarial examples across deep neu- ral network (DNN) models is the crux of a spectrum of black-box attacks. In this paper, we propose a novel method to enhance the black-box trans- ferability of baseline adversarial examples. By establishing a linear map- ping of the intermediate-level discrepancies (between a set of adversarial inputs and their benign counterparts) for predicting the evoked adver- sarial loss, we aim to take full advantage of the optimization procedure of multi-step baseline attacks. We conducted extensive experiments to ver- ify the effectiveness of our method on CIFAR-100 and ImageNet. Experi- mental results demonstrate that it outperforms previous state-of-the-arts considerably. Our code is at https://github.com/qizhangli/ila-plus-plus.