Bigram and Unigram Based Text Attack via Adaptive Monotonic Heuristic Search
Authors: Xinghao Yang, Weifeng Liu, James Bailey, Dacheng Tao, Wei Liu706-714
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the effectiveness of BU-MHS on IMDB, AG s News, and Yahoo! Answers text datasets by attacking four popular DNNs models. Results show that our BU-MHS achieves the highest attack success rate by changing the smallest number of words compared with baselines. |
| Researcher Affiliation | Academia | Xinghao Yang1*, Weifeng Liu2, James Bailey3, Dacheng Tao4, Wei Liu1 1School of Computer Science, University of Technology Sydney, Australia, 2School of Information and Control Engineering, China University of Petroleum (East China), China, 3School of Computing and Information Systems, The University of Melbourne, Australia, 4School of Computer Science, Faculty of Engineering, The University of Sydney, Australia |
| Pseudocode | Yes | Algorithm 1: The proposed BU-MHS Attack Algorithm Input: Sample text with n words X = (w1, , wn) Input: Maximum word replacement bond M Input: Classifier F Output: Adversarial example Xadv |
| Open Source Code | Yes | We provide code and data with a fully anonymous link3 to ensure reproducibility. 3https://github.com/Adv Attack/Text Attack |
| Open Datasets | Yes | We conduct experiments on three publicly available benchmarks. IMDB (Maas et al. 2011) is a binary sentiment classification dataset containing 50,000 movie reviews. AG s News (Zhang, Zhao, and Le Cun 2015) is a news classification dataset with 127600 samples belonging to 4 topic classes. Yahoo! Answers (Zhang, Zhao, and Le Cun 2015) is a ten-class topic dataset with 1,400,000 train samples and 60,000 test samples. |
| Dataset Splits | No | The paper mentions '1,400,000 train samples and 60,000 test samples' for Yahoo! Answers and total sizes for other datasets but does not explicitly provide validation split percentages or counts for any dataset used in the experiments. |
| Hardware Specification | Yes | We conduct all experiments on Enterprise Linux Workstation 7.7 with 2.7GHz CPU frequency and 176GB memory. |
| Software Dependencies | No | The paper mentions software like Keras, ADAM optimizer, and Text Attack framework, but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We train all the DNNs models using the ADAM optimizer (Kingma and Ba 2015), where parameters are set as: learning rate = 0.001, β1 = 0.9, β2 = 0.999, ϵ = 10 7. |