IsoAdverse Documentation

Introduction

Welcome to IsoAdverse, a Python package designed to simulate adversarial on pre-trained language models (pre-LLM models). This package implements a range of attacks as described in recent research to help secure your AI Agents and LLMs.

Installation

To install the IsoAdverse package, you can use pip:

pip install iso-adverse

Quickstart

Here’s a quick example of how to use IsoAdverse to train a BERT model with adversarial training:

import torch
from isoadverse.utils.data_loader import get_data_loader
from isoadverse.utils.model_loader import get_model_and_tokenizer

# Load data and model
texts = ["This is a positive sentence.", "This is a negative sentence."]
labels = torch.tensor([1, 0])
train_loader = get_data_loader(texts, labels, batch_size=2)

model, tokenizer = get_model_and_tokenizer(model_name='bert-base-uncased')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)

Attacks

IsoAdverse implements several adversarial attacks on text data. Below are the details of each attack.

Fast Gradient Sign Method (FGSM)

The FGSM attack perturbs the input text by leveraging the gradients of the loss with respect to the input.

from isoadverse.attacks.text_fgsm import text_fgsm_attack
print("Running FGSM Attack...")
perturbed_text = text_fgsm_attack(model, tokenizer, texts[0], torch.tensor([labels[0]]), epsilon=0.3)
print("Original Text:", texts[0])
print("Perturbed Text:", tokenizer.decode(perturbed_text[0]))

Projected Gradient Descent (PGD)

The PGD attack is an iterative attack method that performs multiple steps of FGSM.

from isoadverse.attacks.text_pgd import text_pgd_attack

print("\nRunning PGD Attack...")
perturbed_ids = text_fgsm_attack(model, tokenizer, texts[0], torch.tensor([labels[0]]), epsilon=0.3)
print("Original Text:", texts[0])
print("Perturbed Text:", tokenizer.decode(perturbed_ids[0]))

TextBugger

TextBugger perturbs the text by introducing character-level changes.

from isoadverse.attacks.textbugger import textbugger_attack

print("\nRunning TextBugger Attack...")
perturbed_text = textbugger_attack(texts[0], num_bugs=5)
print("Original Text:", texts[0])
print("Perturbed Text:", perturbed_text)

DeepWordBug

DeepWordBug introduces word-level perturbations by modifying words in the text.

from isoadverse.attacks.deepwordbug import deepwordbug_attack

print("\nRunning DeepWordBug Attack...")
perturbed_text = deepwordbug_attack(texts[0], num_bugs=5)
print("Original Text:", texts[0])
print("Perturbed Text:", perturbed_text)

Utilities

IsoAdverse includes utility functions for loading data and models, making it easier to integrate into your existing workflow.

Data Loader

The data loader utility helps load and prepare text datasets for training and evaluation.

from isoadverse.utils.data_loader import get_data_loader

train_loader = get_data_loader(texts, labels, batch_size=2)

Model Loader

The model loader utility provides pre-trained models and tokenizers.

from isoadverse.utils.model_loader import get_model_and_tokenizer

model, tokenizer = get_model_and_tokenizer(model_name='bert-base-uncased')