SummAttack

SummAttack is an open-source framework designed for conducting adversarial attacks on large language models specifically tailored for the summarization task.

Running the pipeline

To run the pipeline, you need to fill the config file (examples in the configs directory) with chosen attacks, metrics and parameters. Then, you can run the pipeline with the following command:

python3 -m carl.carl_runner --experiment <path_to_config_file>

Attacks

To add your own attack you need to add inherit from the Attack class and implement the attack method. Then, you need to add your attack in the attacks/__init__.py file.

Each attack should have a unique name. In the attack function you should return the adversarial example and the number of changes made to the input. The adversarial example should be a string with the same number of sentences as the input. The number of changes should be an integer.

class Attack(ABC):
    def __init__(self):
        self.name = None

    def attack(self, sentences: List[str]) -> Tuple[str, int]:
        changes = 0
        return ' '.join(sentences), changes

    def __call__(self, sentences: List[str]) -> str:
        return self.attack(sentences)

    def __name__(self):
        return self.name
        

Metrics

To add your own metric you need to add inherit from the Metric class and implement the compute method. Then, you need to add your metric in the metrics/__init__.py file.

Each metric should have a unique name. In the compute function you should return the value of the metric. The value should be a float.

There are two types of metrics: MetricSummarytoSummary and MetricOriginalTextToSummary. The first type of metric takes two summaries as input and the second type takes a summary and the original text as input.

class MetricSummarytoSummary(ABC):
    def __init__(self):
        self.name = None
        self.type = None

    def compute(self, generated_summary: str, reference_summary: str) -> float:
        pass

    def __call__(self, generated_summary: str, reference_summary: str) -> float:
        return self.compute(generated_summary, reference_summary)

    def __name__(self):
        return self.name


class MetricOriginalTextToSummary(ABC):
    def __init__(self):
        self.name = None
        self.type = None

    def compute(self, generated_summary: str, original_text: str) -> float:
        pass

    def __call__(self, generated_summary: str, original_text: str) -> float:
        return self.compute(generated_summary, original_text)

    def __name__(self):
        return self.name

Summarizers

The most popular commercial APIs (OpenAI) and open-source libraries (HuggingFace) are supported. To add your own summarizer you need to add inherit from the Summarizer class and implement the summarize method. Then, you need to add your summarizer in the summarizers/__init__.py file.

class Summarizer(ABC):
    def __init__(self):
        self.name = None

    def generate_summary(self, text: str) -> str:
        pass

    def __call__(self, text: str) -> str:
        return self.summarize(text)

    def __name__(self):
        return self.name