File size: 21,418 Bytes
4943752 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 |
"""
Attack Class
============
"""
from collections import OrderedDict
from typing import List, Union
import lru
import torch
import textattack
from textattack.attack_results import (
FailedAttackResult,
MaximizedAttackResult,
SkippedAttackResult,
SuccessfulAttackResult,
)
from textattack.constraints import Constraint, PreTransformationConstraint
from textattack.goal_function_results import GoalFunctionResultStatus
from textattack.goal_functions import GoalFunction
from textattack.models.wrappers import ModelWrapper
from textattack.search_methods import SearchMethod
from textattack.shared import AttackedText, utils
from textattack.transformations import CompositeTransformation, Transformation
class Attack:
"""An attack generates adversarial examples on text.
An attack is comprised of a goal function, constraints, transformation, and a search method. Use :meth:`attack` method to attack one sample at a time.
Args:
goal_function (:class:`~textattack.goal_functions.GoalFunction`):
A function for determining how well a perturbation is doing at achieving the attack's goal.
constraints (list of :class:`~textattack.constraints.Constraint` or :class:`~textattack.constraints.PreTransformationConstraint`):
A list of constraints to add to the attack, defining which perturbations are valid.
transformation (:class:`~textattack.transformations.Transformation`):
The transformation applied at each step of the attack.
search_method (:class:`~textattack.search_methods.SearchMethod`):
The method for exploring the search space of possible perturbations
transformation_cache_size (:obj:`int`, `optional`, defaults to :obj:`2**15`):
The number of items to keep in the transformations cache
constraint_cache_size (:obj:`int`, `optional`, defaults to :obj:`2**15`):
The number of items to keep in the constraints cache
Example::
>>> import textattack
>>> import transformers
>>> # Load model, tokenizer, and model_wrapper
>>> model = transformers.AutoModelForSequenceClassification.from_pretrained("textattack/bert-base-uncased-imdb")
>>> tokenizer = transformers.AutoTokenizer.from_pretrained("textattack/bert-base-uncased-imdb")
>>> model_wrapper = textattack.models.wrappers.HuggingFaceModelWrapper(model, tokenizer)
>>> # Construct our four components for `Attack`
>>> from textattack.constraints.pre_transformation import RepeatModification, StopwordModification
>>> from textattack.constraints.semantics import WordEmbeddingDistance
>>> goal_function = textattack.goal_functions.UntargetedClassification(model_wrapper)
>>> constraints = [
... RepeatModification(),
... StopwordModification()
... WordEmbeddingDistance(min_cos_sim=0.9)
... ]
>>> transformation = WordSwapEmbedding(max_candidates=50)
>>> search_method = GreedyWordSwapWIR(wir_method="delete")
>>> # Construct the actual attack
>>> attack = Attack(goal_function, constraints, transformation, search_method)
>>> input_text = "I really enjoyed the new movie that came out last month."
>>> label = 1 #Positive
>>> attack_result = attack.attack(input_text, label)
"""
def __init__(
self,
goal_function: GoalFunction,
constraints: List[Union[Constraint, PreTransformationConstraint]],
transformation: Transformation,
search_method: SearchMethod,
transformation_cache_size=2**15,
constraint_cache_size=2**15,
):
"""Initialize an attack object.
Attacks can be run multiple times.
"""
assert isinstance(
goal_function, GoalFunction
), f"`goal_function` must be of type `textattack.goal_functions.GoalFunction`, but got type `{type(goal_function)}`."
assert isinstance(
constraints, list
), "`constraints` must be a list of `textattack.constraints.Constraint` or `textattack.constraints.PreTransformationConstraint`."
for c in constraints:
assert isinstance(
c, (Constraint, PreTransformationConstraint)
), "`constraints` must be a list of `textattack.constraints.Constraint` or `textattack.constraints.PreTransformationConstraint`."
assert isinstance(
transformation, Transformation
), f"`transformation` must be of type `textattack.transformations.Transformation`, but got type `{type(transformation)}`."
assert isinstance(
search_method, SearchMethod
), f"`search_method` must be of type `textattack.search_methods.SearchMethod`, but got type `{type(search_method)}`."
self.goal_function = goal_function
self.search_method = search_method
self.transformation = transformation
self.is_black_box = (
getattr(transformation, "is_black_box", True) and search_method.is_black_box
)
if not self.search_method.check_transformation_compatibility(
self.transformation
):
raise ValueError(
f"SearchMethod {self.search_method} incompatible with transformation {self.transformation}"
)
self.constraints = []
self.pre_transformation_constraints = []
for constraint in constraints:
if isinstance(
constraint,
textattack.constraints.PreTransformationConstraint,
):
self.pre_transformation_constraints.append(constraint)
else:
self.constraints.append(constraint)
# Check if we can use transformation cache for our transformation.
if not self.transformation.deterministic:
self.use_transformation_cache = False
elif isinstance(self.transformation, CompositeTransformation):
self.use_transformation_cache = True
for t in self.transformation.transformations:
if not t.deterministic:
self.use_transformation_cache = False
break
else:
self.use_transformation_cache = True
self.transformation_cache_size = transformation_cache_size
self.transformation_cache = lru.LRU(transformation_cache_size)
self.constraint_cache_size = constraint_cache_size
self.constraints_cache = lru.LRU(constraint_cache_size)
# Give search method access to functions for getting transformations and evaluating them
self.search_method.get_transformations = self.get_transformations
# Give search method access to self.goal_function for model query count, etc.
self.search_method.goal_function = self.goal_function
# The search method only needs access to the first argument. The second is only used
# by the attack class when checking whether to skip the sample
self.search_method.get_goal_results = self.goal_function.get_results
# Give search method access to get indices which need to be ordered / searched
self.search_method.get_indices_to_order = self.get_indices_to_order
self.search_method.filter_transformations = self.filter_transformations
def clear_cache(self, recursive=True):
self.constraints_cache.clear()
if self.use_transformation_cache:
self.transformation_cache.clear()
if recursive:
self.goal_function.clear_cache()
for constraint in self.constraints:
if hasattr(constraint, "clear_cache"):
constraint.clear_cache()
def cpu_(self):
"""Move any `torch.nn.Module` models that are part of Attack to CPU."""
visited = set()
def to_cpu(obj):
visited.add(id(obj))
if isinstance(obj, torch.nn.Module):
obj.cpu()
elif isinstance(
obj,
(
Attack,
GoalFunction,
Transformation,
SearchMethod,
Constraint,
PreTransformationConstraint,
ModelWrapper,
),
):
for key in obj.__dict__:
s_obj = obj.__dict__[key]
if id(s_obj) not in visited:
to_cpu(s_obj)
elif isinstance(obj, (list, tuple)):
for item in obj:
if id(item) not in visited and isinstance(
item, (Transformation, Constraint, PreTransformationConstraint)
):
to_cpu(item)
to_cpu(self)
def cuda_(self):
"""Move any `torch.nn.Module` models that are part of Attack to GPU."""
visited = set()
def to_cuda(obj):
visited.add(id(obj))
if isinstance(obj, torch.nn.Module):
obj.to(textattack.shared.utils.device)
elif isinstance(
obj,
(
Attack,
GoalFunction,
Transformation,
SearchMethod,
Constraint,
PreTransformationConstraint,
ModelWrapper,
),
):
for key in obj.__dict__:
s_obj = obj.__dict__[key]
if id(s_obj) not in visited:
to_cuda(s_obj)
elif isinstance(obj, (list, tuple)):
for item in obj:
if id(item) not in visited and isinstance(
item, (Transformation, Constraint, PreTransformationConstraint)
):
to_cuda(item)
to_cuda(self)
def get_indices_to_order(self, current_text, **kwargs):
"""Applies ``pre_transformation_constraints`` to ``text`` to get all
the indices that can be used to search and order.
Args:
current_text: The current ``AttackedText`` for which we need to find indices are eligible to be ordered.
Returns:
The length and the filtered list of indices which search methods can use to search/order.
"""
indices_to_order = self.transformation(
current_text,
pre_transformation_constraints=self.pre_transformation_constraints,
return_indices=True,
**kwargs,
)
len_text = len(indices_to_order)
# Convert indices_to_order to list for easier shuffling later
return len_text, list(indices_to_order)
def _get_transformations_uncached(self, current_text, original_text=None, **kwargs):
"""Applies ``self.transformation`` to ``text``, then filters the list
of possible transformations through the applicable constraints.
Args:
current_text: The current ``AttackedText`` on which to perform the transformations.
original_text: The original ``AttackedText`` from which the attack started.
Returns:
A filtered list of transformations where each transformation matches the constraints
"""
transformed_texts = self.transformation(
current_text,
pre_transformation_constraints=self.pre_transformation_constraints,
**kwargs,
)
return transformed_texts
def get_transformations(self, current_text, original_text=None, **kwargs):
"""Applies ``self.transformation`` to ``text``, then filters the list
of possible transformations through the applicable constraints.
Args:
current_text: The current ``AttackedText`` on which to perform the transformations.
original_text: The original ``AttackedText`` from which the attack started.
Returns:
A filtered list of transformations where each transformation matches the constraints
"""
if not self.transformation:
raise RuntimeError(
"Cannot call `get_transformations` without a transformation."
)
if self.use_transformation_cache:
cache_key = tuple([current_text] + sorted(kwargs.items()))
if utils.hashable(cache_key) and cache_key in self.transformation_cache:
# promote transformed_text to the top of the LRU cache
self.transformation_cache[cache_key] = self.transformation_cache[
cache_key
]
transformed_texts = list(self.transformation_cache[cache_key])
else:
transformed_texts = self._get_transformations_uncached(
current_text, original_text, **kwargs
)
if utils.hashable(cache_key):
self.transformation_cache[cache_key] = tuple(transformed_texts)
else:
transformed_texts = self._get_transformations_uncached(
current_text, original_text, **kwargs
)
return self.filter_transformations(
transformed_texts, current_text, original_text
)
def _filter_transformations_uncached(
self, transformed_texts, current_text, original_text=None
):
"""Filters a list of potential transformed texts based on
``self.constraints``
Args:
transformed_texts: A list of candidate transformed ``AttackedText`` to filter.
current_text: The current ``AttackedText`` on which the transformation was applied.
original_text: The original ``AttackedText`` from which the attack started.
"""
filtered_texts = transformed_texts[:]
for C in self.constraints:
if len(filtered_texts) == 0:
break
if C.compare_against_original:
if not original_text:
raise ValueError(
f"Missing `original_text` argument when constraint {type(C)} is set to compare against `original_text`"
)
filtered_texts = C.call_many(filtered_texts, original_text)
else:
filtered_texts = C.call_many(filtered_texts, current_text)
# Default to false for all original transformations.
for original_transformed_text in transformed_texts:
self.constraints_cache[(current_text, original_transformed_text)] = False
# Set unfiltered transformations to True in the cache.
for filtered_text in filtered_texts:
self.constraints_cache[(current_text, filtered_text)] = True
return filtered_texts
def filter_transformations(
self, transformed_texts, current_text, original_text=None
):
"""Filters a list of potential transformed texts based on
``self.constraints`` Utilizes an LRU cache to attempt to avoid
recomputing common transformations.
Args:
transformed_texts: A list of candidate transformed ``AttackedText`` to filter.
current_text: The current ``AttackedText`` on which the transformation was applied.
original_text: The original ``AttackedText`` from which the attack started.
"""
# Remove any occurences of current_text in transformed_texts
transformed_texts = [
t for t in transformed_texts if t.text != current_text.text
]
# Populate cache with transformed_texts
uncached_texts = []
filtered_texts = []
for transformed_text in transformed_texts:
if (current_text, transformed_text) not in self.constraints_cache:
uncached_texts.append(transformed_text)
else:
# promote transformed_text to the top of the LRU cache
self.constraints_cache[
(current_text, transformed_text)
] = self.constraints_cache[(current_text, transformed_text)]
if self.constraints_cache[(current_text, transformed_text)]:
filtered_texts.append(transformed_text)
filtered_texts += self._filter_transformations_uncached(
uncached_texts, current_text, original_text=original_text
)
# Sort transformations to ensure order is preserved between runs
filtered_texts.sort(key=lambda t: t.text)
return filtered_texts
def _attack(self, initial_result):
"""Calls the ``SearchMethod`` to perturb the ``AttackedText`` stored in
``initial_result``.
Args:
initial_result: The initial ``GoalFunctionResult`` from which to perturb.
Returns:
A ``SuccessfulAttackResult``, ``FailedAttackResult``,
or ``MaximizedAttackResult``.
"""
final_result = self.search_method(initial_result)
self.clear_cache()
if final_result.goal_status == GoalFunctionResultStatus.SUCCEEDED:
result = SuccessfulAttackResult(
initial_result,
final_result,
)
elif final_result.goal_status == GoalFunctionResultStatus.SEARCHING:
result = FailedAttackResult(
initial_result,
final_result,
)
elif final_result.goal_status == GoalFunctionResultStatus.MAXIMIZING:
result = MaximizedAttackResult(
initial_result,
final_result,
)
else:
raise ValueError(f"Unrecognized goal status {final_result.goal_status}")
return result
def attack(self, example, ground_truth_output):
"""Attack a single example.
Args:
example (:obj:`str`, :obj:`OrderedDict[str, str]` or :class:`~textattack.shared.AttackedText`):
Example to attack. It can be a single string or an `OrderedDict` where
keys represent the input fields (e.g. "premise", "hypothesis") and the values are the actual input textx.
Also accepts :class:`~textattack.shared.AttackedText` that wraps around the input.
ground_truth_output(:obj:`int`, :obj:`float` or :obj:`str`):
Ground truth output of `example`.
For classification tasks, it should be an integer representing the ground truth label.
For regression tasks (e.g. STS), it should be the target value.
For seq2seq tasks (e.g. translation), it should be the target string.
Returns:
:class:`~textattack.attack_results.AttackResult` that represents the result of the attack.
"""
assert isinstance(
example, (str, OrderedDict, AttackedText)
), "`example` must either be `str`, `collections.OrderedDict`, `textattack.shared.AttackedText`."
if isinstance(example, (str, OrderedDict)):
example = AttackedText(example)
assert isinstance(
ground_truth_output, (int, str)
), "`ground_truth_output` must either be `str` or `int`."
goal_function_result, _ = self.goal_function.init_attack_example(
example, ground_truth_output
)
if goal_function_result.goal_status == GoalFunctionResultStatus.SKIPPED:
return SkippedAttackResult(goal_function_result)
else:
result = self._attack(goal_function_result)
return result
def __repr__(self):
"""Prints attack parameters in a human-readable string.
Inspired by the readability of printing PyTorch nn.Modules:
https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/module.py
"""
main_str = "Attack" + "("
lines = []
lines.append(utils.add_indent(f"(search_method): {self.search_method}", 2))
# self.goal_function
lines.append(utils.add_indent(f"(goal_function): {self.goal_function}", 2))
# self.transformation
lines.append(utils.add_indent(f"(transformation): {self.transformation}", 2))
# self.constraints
constraints_lines = []
constraints = self.constraints + self.pre_transformation_constraints
if len(constraints):
for i, constraint in enumerate(constraints):
constraints_lines.append(utils.add_indent(f"({i}): {constraint}", 2))
constraints_str = utils.add_indent("\n" + "\n".join(constraints_lines), 2)
else:
constraints_str = "None"
lines.append(utils.add_indent(f"(constraints): {constraints_str}", 2))
# self.is_black_box
lines.append(utils.add_indent(f"(is_black_box): {self.is_black_box}", 2))
main_str += "\n " + "\n ".join(lines) + "\n"
main_str += ")"
return main_str
def __getstate__(self):
state = self.__dict__.copy()
state["transformation_cache"] = None
state["constraints_cache"] = None
return state
def __setstate__(self, state):
self.__dict__ = state
self.transformation_cache = lru.LRU(self.transformation_cache_size)
self.constraints_cache = lru.LRU(self.constraint_cache_size)
__str__ = __repr__
|