IMPLEMENTATION OF BLACK BOX ATTACK WITH TRANSFERABLE MODEL BASED EMBEDDING

Zhichao Huang, Tong Zhang The Hong Kong University of Science and Technology

https://openreview.net/forum?id=SJxhNTNYwB

Abstract : "We present a new method for black-box adversarial attack. Unlike previous methods that combined transfer-based and scored-based methods by using the gradient or initialization of a surrogate white-box model, this new method tries to learn a low-dimensional embedding using a pretrained model, and then performs efficient search within the embedding space to attack an unknown target network. The method produces adversarial perturbations with high level semantic patterns that are easily transferable. We show that this approach can greatly improve the query efficiency of black-box adversarial attack across different target network architectures. We evaluate our approach on MNIST, ImageNet and Google Cloud Vision API, resulting in a significant reduction on the number of queries. We also attack adversarially defended networks on CIFAR10 and ImageNet, where our method not only reduces the number of queries, but also improves the attack success rate."

STATE OF THE ART

White-box attack

Black-box attack

Neural Phrase based-translation model

Meilleur résultats sur les neural phrase based-translation model : Huang et al (2018) https://arxiv.org/pdf/1706.05565.pdf

Segmental neural networks

Segmental neural networks : https://arxiv.org/pdf/1511.06018.pdf

Experimentations

Dataset

IWSLT 2014 German to English dataset: https://wit3.fbk.eu/mt.php?release=2014-01 Aller à la ligne German et colonne English.

IWSLT 2015 English to Vietnamese Dataset : https://wit3.fbk.eu/mt.php?release=2015-01 Aller à la ligne english et colonne vietnamien.

Une fois télécharger mettre dans un dossier iwslt2014_2015/

Baseline (à réflechir si on implémente ou non):

NPMT : https://github.com/posenhuang/NPMT , Huang et al (2018) https://arxiv.org/pdf/1706.05565.pdf Seq2Seq attention : https://github.com/pytorch/fairseq , Bahdhanau et al (2014) https://arxiv.org/pdf/1409.0473.pdf Transformers : https://arxiv.org/pdf/1706.03762.pdf

Metrique d'évaluations :

BLEU. article https://www.aclweb.org/anthology/P02-1040.pdf BLEU. Code pytorch https://pytorchnlp.readthedocs.io/en/latest/_modules/torchnlp/metrics/bleu.html