Attacks on Chinese Dataset

In this example, we show how to use OpenAttack to attack your classifier on a Chinese dataset.

Load a Chinese Dataset

In this example, we read in the Chinese part of the amazon_reviews_multi dataset using datasets package.

1
2
3
4
5
6
def dataset_mapping(x):
    return {
        "x": x["review_body"],
        "y": x["stars"],
    }
dataset = datasets.load_dataset("amazon_reviews_multi",'zh',split="train[:5]").map(dataset_mapping)

We selected the first five examples in the dataset and used the dataset_mapping function to map the review_body field to x and the stars field to y which is the field victim models need to predict.

Load a Chinese Victim Model

OpenAttack provides a trained Chinese BERT classification model on the amazon_reviews_multi dataset.

In this example, we use OpenAttack.loadVictim to load the victim model and move it to a GPU.

1
victim = OpenAttack.loadVictim("BERT.AMAZON_ZH").to("cuda:0")

Initialize an Attacker

Almost all Attacker take a lang argument which means the language you use. We need to specify the lang argument before we start the attack.

1
attacker = OpenAttack.attackers.PWWSAttacker(lang="chinese")

In this example, we chose the PWWSAttacker and set lang to “chinese”.

Use a Chinese AttackEval

The OpenAttack toolkit supports the lang option in AttackEval class to perform attack evaluation in different languages.

Just like the other examples, we only need one simple line of code to start the evaluation.

1
OpenAttack.AttackEval(attacker, victim, lang="chinese").eval(dataset, progress_bar=True)

Complete Code

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
import OpenAttack
import datasets
def main():
    attacker = OpenAttack.attackers.PWWSAttacker(lang="chinese")
    victim = OpenAttack.loadVictim("BERT.AMAZON_ZH").to("cuda:0")
    def dataset_mapping(x):
        return {
            "x": x["review_body"],
            "y": x["stars"],
        }
    dataset = datasets.load_dataset("amazon_reviews_multi",'zh',split="train[:20]").map(function=dataset_mapping)
    attack_eval = OpenAttack.AttackEval(attacker, victim)
    attack_eval.eval(dataset, visualize=True, progress_bar=True)