We determine the usefulness of our toolkit as a counter-measure against character-level adversarial
attacks (§ 5.2). We find that our models are better defenses to
adversarial attacks than previously proposed spell checkers. In this section, we provide a guide to utilize one of the
spell checkers in our toolkit- SC-LSTM plus ELMO (at input) (aka. ElmosclstmChecker) in the setup
of Adversarial-Misspellings (Pruthi et al. 2019)
.
git clone https://github.com/danishpruthi/Adversarial-Misspellings; cd Adversarial-Misspellings
mv ../Adversarial-Misspellings-arxiv/main.py ./And then the following commands can be utilized
- Word level attacks
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-word-only --attack <swap,drop or key> --num-attacks 2 --model bilstm-word --defense --sc-elmoscrnn --backoff <neutral or pass-through>
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-word-only --attack <add or all> --num-attacks 2 --model bilstm-word --defense --sc-elmoscrnn --small --backoff <neutral or pass-through>- Char level attacks
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-char-only --attack <swap,drop or key> --num-attacks 2 --model bilstm-char --defense --sc-elmoscrnn --backoff <neutral or pass-through>
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-char-only --attack <add or all> --num-attacks 2 --model bilstm-char --defense --sc-elmoscrnn --small --backoff <neutral or pass-through>- Char+Word level attacks
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-word-plus-char --attack <swap,drop or key> --num-attacks 2 --model bilstm --defense --sc-elmoscrnn --backoff <neutral or pass-through>
CUDA_VISIBLE_DEVICES=x python main.py --dynet-seed 1 --mode dev --load model_dumps/bilstm-word-plus-char --attack <add or all> --num-attacks 2 --model bilstm --defense --sc-elmoscrnn --small --backoff <neutral or pass-through>