Spaces:
Sleeping
Sleeping
File size: 15,276 Bytes
957e2dc |
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 |
import torch
import torch.nn as nn
from typing import Tuple, Union
from torch.utils.data import Dataset, TensorDataset, DataLoader
from src.utils.writer import Writer
from src.pipelines.pipeline import Pipeline
from src.loss.adversarial import AdversarialLoss
from src.loss.auxiliary import AuxiliaryLoss
################################################################################
# Base class for offline adversarial attacks
################################################################################
class OfflineAttack:
def __init__(self,
pipeline: Pipeline,
adv_loss: AdversarialLoss,
aux_loss: AuxiliaryLoss = None,
batch_size: int = 32,
rand_evals: int = 0,
writer: Writer = None,
**kwargs
):
"""
Base class for offline attacks. Subclasses must override the
`evaluate_batch()` method.
Offline attacks optimize perturbations of benign inputs without
real-time performance constraints. Optimization is performed using a
stored Pipeline object, encompassing a victim model, acoustic
simulation, and adversarial defenses.
:param pipeline: a Pipeline object wrapping a (defended) classifier
:param adv_loss: AdversarialLoss object encapsulating attacker objective
:param aux_loss: optional AuxiliaryLoss object encapsulating
some perceptibility objective
:param batch_size: batch size for attack
:param rand_evals: randomly-resampled simulated evaluations per each
final generated attack
"""
self.pipeline = pipeline
# ensure gradients flow through PyTorch RNN layers
self._pipeline_rnn_grad()
self.adv_loss = adv_loss
self.aux_loss = aux_loss
self.batch_size = batch_size
self.rand_evals = rand_evals
# log attack progress
self.writer = writer
# track batch inputs
self._batch_id = 0
self._iter_id = 0
# optional data-loading arguments
self.pin_memory = kwargs.get('pin_memory', False)
self.num_workers = kwargs.get('num_workers', 0)
self._check_loss()
def _pipeline_rnn_grad(self):
"""
PyTorch requires any recurrent modules be placed in `train` mode to
enable backpropagation through the pipeline.
"""
for m in self.pipeline.modules():
if isinstance(m, nn.RNNBase):
m.train()
def _check_loss(self):
"""
Validate adversarial and auxiliary losses
"""
assert self.adv_loss is not None, 'Must provide adversarial loss'
assert self.adv_loss.reduction in ['none', None], \
'All losses must provide unreduced scores'
assert self.aux_loss is None or \
self.aux_loss.reduction in ['none', None], \
'All losses must provide unreduced scores'
@staticmethod
def _create_dataset(x: torch.Tensor, y: torch.Tensor):
"""
If attack inputs are given as tensors, create a simple dataset
"""
# require batch dimension
assert x.ndim >= 2
dataset = TensorDataset(
x.type(torch.float32),
y.type(torch.float32),
)
return dataset
def _compute_detection_array(self, x_adv, *args, **kwargs):
"""
Pass attack audio through any detection defenses in stored Pipeline, and
return boolean detection flags for each input
"""
flags, scores = self.pipeline.detect(x_adv)
return flags
@torch.no_grad()
def _compute_success_array(self,
x: torch.Tensor,
y: torch.Tensor,
x_adv: torch.Tensor,
*args,
**kwargs
):
"""
Pass attack audio through stored Pipeline and determine adversarial
success for each input
"""
# obtain 'clean' and adversarial predictions of stored Pipeline
preds = self.pipeline(x.detach())
adv_preds = self.pipeline(x_adv.detach())
# for a targeted attack, attempt to match given targets
if self.adv_loss.targeted:
attack_success = self.pipeline.match_predict(adv_preds, y)
# for an untargeted attack, attempt to evade clean predictions
else:
attack_success = ~self.pipeline.match_predict(adv_preds, preds)
return attack_success
def _log_step(self,
x: torch.Tensor,
x_adv: torch.Tensor,
y: torch.Tensor,
adv_loss: Union[float, torch.Tensor] = None,
det_loss: Union[float, torch.Tensor] = None,
aux_loss: Union[float, torch.Tensor] = None,
success_rate: Union[float, torch.Tensor] = None,
detection_rate: Union[float, torch.Tensor] = None,
idx: int = 0,
tag: str = None,
*args,
**kwargs
):
"""
Log attack progress.
:param x: batch of original inputs
:param x_adv: batch of adversarial inputs
:param y: batch of adversarial targets
:param adv_loss: adversarial loss value
:param det_loss: detection loss value
:param aux_loss: auxiliary loss value
:param success_rate: adversarial success rate
:param detection_rate: adversarial defense detection rate
:param idx: batch index for logging individual examples
"""
if self.writer is None or self._iter_id % self.writer.log_iter:
return
if tag is None:
tag = f'{self.__class__.__name__}-batch-{self._batch_id}'
x = x.clone().detach()
x_adv = x_adv.clone().detach()
# compute losses and simulated audio
with torch.no_grad():
outputs_adv = self.pipeline(x_adv)
simulated = self.pipeline.simulate(x)
simulated_adv = self.pipeline.simulate(x_adv)
# if adversarial loss is not provided, compute
if adv_loss is None:
adv_loss = self.adv_loss(outputs_adv, y).mean()
# if detector loss or rate is not provided, compute
if det_loss is None or detection_rate is None:
flags, scores = self.pipeline.detect(x_adv)
det_loss = scores.mean()
detection_rate = torch.mean(1.0 * flags)
# if auxiliary loss is not provided, compute
if aux_loss is None:
aux_loss = 0.0 if self.aux_loss is None else self.aux_loss(
x_adv, x
).mean()
# if adversarial success rate is not provided, compute
if success_rate is None:
success = self._compute_success_array(
x=x,
x_adv=x_adv,
y=y
)
success_rate = torch.mean(1.0 * success)
# unperturbed input
self.writer.log_audio(
x[idx],
f"{tag}/original",
global_step=self._iter_id
)
# simulated unperturbed input
self.writer.log_audio(
simulated[idx],
f"{tag}/simulated-original",
global_step=self._iter_id
)
# adversarial input
self.writer.log_audio(
x_adv[idx],
f"{tag}/adversarial",
global_step=self._iter_id
)
# simulated adversarial input
self.writer.log_audio(
simulated_adv[idx],
f"{tag}/simulated-adversarial",
global_step=self._iter_id
)
# adversarial loss value
self.writer.log_scalar(
adv_loss,
f"{tag}/adversarial-loss",
global_step=self._iter_id
)
# detector loss value
self.writer.log_scalar(
det_loss,
f"{tag}/detector-loss",
global_step=self._iter_id
)
# auxiliary loss value
self.writer.log_scalar(
aux_loss,
f"{tag}/auxiliary-loss",
global_step=self._iter_id
)
# adversarial success rate
self.writer.log_scalar(
success_rate,
f"{tag}/success-rate",
global_step=self._iter_id
)
# adversarial detection rate
self.writer.log_scalar(
detection_rate,
f"{tag}/detection-rate",
global_step=self._iter_id
)
def _evaluate_batch(self,
x: torch.Tensor,
y: torch.Tensor,
**kwargs
):
"""
Perform attack on a batch of inputs.
:param x: input tensor of shape (n_batch, ...)
:param y: targets tensor of shape (n_batch, ...) in case of targeted
attack; original labels tensor of shape (n_batch, ...) in
case of untargeted attack
"""
raise NotImplementedError()
@torch.no_grad()
def evaluate(self,
x: torch.Tensor = None,
y: torch.Tensor = None,
dataset: Dataset = None,
**kwargs
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Perform attack given input-target pairs, optionally in the form of a
Dataset object. Random evaluations will then be conducted on all
generated attacks.
:param x: audio input, shape (n_batch, ..., signal_length)
:param y: targets, shape (n_batch, ...)
:param dataset: optionally, provide inputs and targets as dataset
:return: tuple holding
* adversarial audio (n_batch, ..., signal_length)
* boolean adversarial success indicators (n_batch,)
* boolean adversarial detection indicators (n_batch,)
"""
assert (x is not None and y is not None) or dataset is not None
# prepare batched data-loading, store original device
if dataset is None:
orig_device = x.device
dataset = self._create_dataset(x, y)
x_ref = x[0:1].clone().detach()
else:
ref_batch = next(iter(dataset))
if isinstance(ref_batch, tuple):
x_ref = ref_batch[0]
elif isinstance(ref_batch, dict):
x_ref = ref_batch['x']
else:
x_ref = ref_batch
orig_device = x_ref.device
# prepare to compute attack success and detection rates
attack_success = torch.zeros(
len(dataset), dtype=torch.float).to(self.pipeline.device)
attack_detection = torch.zeros(
len(dataset), dtype=torch.float).to(self.pipeline.device)
# prepare to store attack outputs
adv_x = torch.stack(
[torch.zeros(x_ref.shape)] * len(dataset),
dim=0
).to(self.pipeline.device)
data_loader = DataLoader(
dataset=dataset,
batch_size=self.batch_size,
shuffle=False,
drop_last=False,
num_workers=self.num_workers,
pin_memory=self.pin_memory
)
# compute attacks with batching
for (batch_id, batch_all) in enumerate(data_loader):
self._batch_id = batch_id
# allow for different dataset formats
if isinstance(batch_all, tuple):
batch_all = {
'x': batch_all[0],
'y': batch_all[1]
}
# match devices
for k in batch_all.keys():
batch_all[k] = batch_all[k].to(self.pipeline.device)
batch_index_1 = batch_id * self.batch_size
batch_index_2 = (batch_id + 1) * self.batch_size
# compute attacks for given batch
adversarial_batch = self._evaluate_batch(
**batch_all, **kwargs
)
# if no random trials, evaluate once
if not self.rand_evals:
# compute and store success rates for batch
attack_success_batch = self._compute_success_array(
**batch_all,
x_adv=adversarial_batch
).reshape(-1).type(torch.float32)
# compute and store detection rates for batch
attack_detection_batch = self._compute_detection_array(
x_adv=adversarial_batch
).reshape(-1).type(torch.float32)
# otherwise, perform multiple random evaluations per attack
else:
# track batch success and detection over random evaluation
success_combined_batch = []
detection_combined_batch = []
for i in range(self.rand_evals):
# randomly sample simulation parameters
self.pipeline.sample_params()
# compute and store success rates for batch
rand_success_batch = self._compute_success_array(
**batch_all,
x_adv=adversarial_batch
).reshape(-1, 1)
success_combined_batch.append(rand_success_batch)
# compute and store detection rates for batch
rand_detection_batch = self._compute_detection_array(
x_adv=adversarial_batch
)
detection_combined_batch.append(rand_detection_batch)
# average results over all trials
attack_success_batch = (1.0 * torch.cat(
success_combined_batch, dim=-1
)).mean(dim=-1)
attack_detection_batch = (1.0 * torch.cat(
detection_combined_batch, dim=-1
)).mean(dim=-1)
# store generated attack audio
adv_x[batch_index_1:batch_index_2] = adversarial_batch
# store success rate per generated attack
attack_success[batch_index_1:batch_index_2] = attack_success_batch
# store detection rate per generated attack
attack_detection[batch_index_1:batch_index_2] = attack_detection_batch
return (adv_x.to(orig_device),
attack_success.to(orig_device),
attack_detection.to(orig_device))
|