/limeattack

LimeAttack@AAAI24

Primary LanguageJavaScript

AAAI24@LimeAttack: Local Explainable Method for Textual Hard-Label Adversarial Attack

LimeAttack's code:

Requirements

  • py3.10
  • boto3==1.26.28
  • botocore==1.29.28
  • torch == 1.12.1+cu116
  • tensorflow-gpu == 2.11.0（optional）
  • tensorflow-hub == 0.12.0（optional）
  • numpy == 1.23.2
  • nltk == 3.7
  • scipy == 1.9.1

Datesets and Victim Model

There are MR, SST-2 , AG, Yahoo and SNLI, MNLI and MNLIm datasets. We adopt the pretrained models provided including BERT,CNN,LSTM. These data and models are adopted from HLBB or TextAttack.

Dependencies

glove.6B.200d.txt and counter-fitted-vectors.txt can be obtained from TextFooler

File Description

  • LimeAttack_classification.py: Attack the victim model for text classification with LimeAttack.

run

bash attack_mr.sh