homoglyph-attack.py

import random
import verifiers as vf
from confusable_homoglyphs import confusables
from datasets import Dataset


def corrupt_text(text: str, max_replacement_rate: float = 0.02) -> str:
    """Corrupt the text by replacing characters with their homoglyphs."""
    nonspace_chars = [char for char in text if not char.isspace()]
    confusable_chars = [
        char for char in nonspace_chars if char in confusables.confusables_data
    ]
    random.shuffle(confusable_chars)
    if not confusable_chars:
        return text

    max_replacements = int(len(text) * max_replacement_rate)
    for _ in range(max_replacements):
        char = confusable_chars.pop()
        indices = [i for i, c in enumerate(text) if c == char]
        index = random.choice(indices)
        text = (
            text[:index]
            + random.choice(confusables.confusables_data[char])["c"]
            + text[(index + 1) :]
        )
    return text


def corrupt_example(example, column, max_replacement_rate):
    example[column] = corrupt_text(example[column], max_replacement_rate)
    return example

def corrupt_column(dataset: Dataset, column: str, max_replacement_rate=0.02):
    """Corrupt the specified column of a dataset by replacing characters with their homoglyphs."""
    return dataset.map(lambda example: corrupt_example(example, column, max_replacement_rate))


def corrupt_environment(
    env: vf.Environment, input_column: str = "text", max_replacement_rate: float = 0.02
) -> vf.Environment:
    """
    Loads a custom environment.
    """
    env.dataset = corrupt_column(
        env.dataset, input_column, max_replacement_rate=max_replacement_rate
    )
    return env


if __name__ == "__main__":
    # Example usage
    text = "Hello, world!"
    corrupted = corrupt_text(text)
    print(corrupted)