TextHoaxer: Budgeted Hard-Label Adversarial Attacks on Text
Authors: Muchao Ye, Chenglin Miao, Ting Wang, Fenglong Ma3877-3884
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on eight text datasets against three representative natural language models, and experimental results show that Text Hoaxer can generate high-quality adversarial examples with higher semantic similarity and lower perturbation rate under the tight-budget setting. |
| Researcher Affiliation | Academia | 1The Pennsylvania State University, University Park, Pennsylvania 16802 2University of Georgia, Athens, Georgia 30602 |
| Pseudocode | No | The paper describes the optimization procedure in narrative form with equations, but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation code is available at https://github.com/machinelearning4health/Text Hoaxer. |
| Open Datasets | Yes | We adopt the following text datasets that are collected against different text classification and natural language inference: (1) MR (Pang and Lee 2005), (2) AG (Zhang, Zhao, and Le Cun 2015), (3) Yahoo (Zhang, Zhao, and Le Cun 2015), (4) Yelp (Zhang, Zhao, and Le Cun 2015), (5) IMDB (Maas et al. 2011), (6) SNLI (Bowman et al. 2015) and (7) MNLI (Williams, Nangia, and Bowman 2018), and (8) m MNLI, a variant of the MNLI dataset with mismatched premise and hypotheses pairs. |
| Dataset Splits | No | The paper mentions "taking the same 1,000 test samples of each dataset" but does not specify the training or validation splits, proportions, or counts for the datasets used. |
| Hardware Specification | Yes | Text Hoaxer is implemented with an NVIDIA V100 GPU. |
| Software Dependencies | No | The paper mentions "Counter-Fitted Word Vectors" and general models like "BERT", "Word CNN", "Word LSTM", but does not list specific software packages or libraries with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We set λ1 = 1, λ2 = λ3 = 0.1, β = 0.5, η1 = 0.3 and η2 = 0.05, and all γi s are initialized as 0.3. |