File size: 14,028 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 |
"""
Particle Swarm Optimization
====================================
Reimplementation of search method from Word-level Textual Adversarial
Attacking as Combinatorial Optimization by Zang et.
al
`<https://www.aclweb.org/anthology/2020.acl-main.540.pdf>`_
`<https://github.com/thunlp/SememePSO-Attack>`_
"""
import copy
import numpy as np
from textattack.goal_function_results import GoalFunctionResultStatus
from textattack.search_methods import PopulationBasedSearch, PopulationMember
from textattack.shared import utils
from textattack.shared.validators import transformation_consists_of_word_swaps
class ParticleSwarmOptimization(PopulationBasedSearch):
"""Attacks a model with word substiutitions using a Particle Swarm
Optimization (PSO) algorithm. Some key hyper-parameters are setup according
to the original paper:
"We adjust PSO on the validation set of SST and set ω_1 as 0.8 and ω_2 as 0.2.
We set the max velocity of the particles V_{max} to 3, which means the changing
probability of the particles ranges from 0.047 (sigmoid(-3)) to 0.953 (sigmoid(3))."
Args:
pop_size (:obj:`int`, optional): The population size. Defaults to 60.
max_iters (:obj:`int`, optional): The maximum number of iterations to use. Defaults to 20.
post_turn_check (:obj:`bool`, optional): If `True`, check if new position reached by moving passes the constraints. Defaults to `True`
max_turn_retries (:obj:`bool`, optional): Maximum number of movement retries if new position after turning fails to pass the constraints.
Applied only when `post_movement_check` is set to `True`.
Setting it to 0 means we immediately take the old position as the new position upon failure.
"""
def __init__(
self, pop_size=60, max_iters=20, post_turn_check=True, max_turn_retries=20
):
self.max_iters = max_iters
self.pop_size = pop_size
self.post_turn_check = post_turn_check
self.max_turn_retries = 20
self._search_over = False
self.omega_1 = 0.8
self.omega_2 = 0.2
self.c1_origin = 0.8
self.c2_origin = 0.2
self.v_max = 3.0
def _perturb(self, pop_member, original_result):
"""Perturb `pop_member` in-place.
Replaces a word at a random in `pop_member` with replacement word that maximizes increase in score.
Args:
pop_member (PopulationMember): The population member being perturbed.
original_result (GoalFunctionResult): Result of original sample being attacked
Returns:
`True` if perturbation occured. `False` if not.
"""
# TODO: Below is very slow and is the main cause behind memory build up + slowness
best_neighbors, prob_list = self._get_best_neighbors(
pop_member.result, original_result
)
random_result = np.random.choice(best_neighbors, 1, p=prob_list)[0]
if random_result == pop_member.result:
return False
else:
pop_member.attacked_text = random_result.attacked_text
pop_member.result = random_result
return True
def _equal(self, a, b):
return -self.v_max if a == b else self.v_max
def _turn(self, source_text, target_text, prob, original_text):
"""
Based on given probabilities, "move" to `target_text` from `source_text`
Args:
source_text (PopulationMember): Text we start from.
target_text (PopulationMember): Text we want to move to.
prob (np.array[float]): Turn probability for each word.
original_text (AttackedText): Original text for constraint check if `self.post_turn_check=True`.
Returns:
New `Position` that we moved to (or if we fail to move, same as `source_text`)
"""
assert len(source_text.words) == len(
target_text.words
), "Word length mismatch for turn operation."
assert len(source_text.words) == len(
prob
), "Length mismatch for words and probability list."
len_x = len(source_text.words)
num_tries = 0
passed_constraints = False
while num_tries < self.max_turn_retries + 1:
indices_to_replace = []
words_to_replace = []
for i in range(len_x):
if np.random.uniform() < prob[i]:
indices_to_replace.append(i)
words_to_replace.append(target_text.words[i])
new_text = source_text.attacked_text.replace_words_at_indices(
indices_to_replace, words_to_replace
)
indices_to_replace = set(indices_to_replace)
new_text.attack_attrs["modified_indices"] = (
source_text.attacked_text.attack_attrs["modified_indices"]
- indices_to_replace
) | (
target_text.attacked_text.attack_attrs["modified_indices"]
& indices_to_replace
)
if "last_transformation" in source_text.attacked_text.attack_attrs:
new_text.attack_attrs[
"last_transformation"
] = source_text.attacked_text.attack_attrs["last_transformation"]
if not self.post_turn_check or (new_text.words == source_text.words):
break
if "last_transformation" in new_text.attack_attrs:
passed_constraints = self._check_constraints(
new_text, source_text.attacked_text, original_text=original_text
)
else:
passed_constraints = True
if passed_constraints:
break
num_tries += 1
if self.post_turn_check and not passed_constraints:
# If we cannot find a turn that passes the constraints, we do not move.
return source_text
else:
return PopulationMember(new_text)
def _get_best_neighbors(self, current_result, original_result):
"""For given current text, find its neighboring texts that yields
maximum improvement (in goal function score) for each word.
Args:
current_result (GoalFunctionResult): `GoalFunctionResult` of current text
original_result (GoalFunctionResult): `GoalFunctionResult` of original text.
Returns:
best_neighbors (list[GoalFunctionResult]): Best neighboring text for each word
prob_list (list[float]): discrete probablity distribution for sampling a neighbor from `best_neighbors`
"""
current_text = current_result.attacked_text
neighbors_list = [[] for _ in range(len(current_text.words))]
transformed_texts = self.get_transformations(
current_text, original_text=original_result.attacked_text
)
for transformed_text in transformed_texts:
diff_idx = next(
iter(transformed_text.attack_attrs["newly_modified_indices"])
)
neighbors_list[diff_idx].append(transformed_text)
best_neighbors = []
score_list = []
for i in range(len(neighbors_list)):
if not neighbors_list[i]:
best_neighbors.append(current_result)
score_list.append(0)
continue
neighbor_results, self._search_over = self.get_goal_results(
neighbors_list[i]
)
if not len(neighbor_results):
best_neighbors.append(current_result)
score_list.append(0)
else:
neighbor_scores = np.array([r.score for r in neighbor_results])
score_diff = neighbor_scores - current_result.score
best_idx = np.argmax(neighbor_scores)
best_neighbors.append(neighbor_results[best_idx])
score_list.append(score_diff[best_idx])
prob_list = normalize(score_list)
return best_neighbors, prob_list
def _initialize_population(self, initial_result, pop_size):
"""
Initialize a population of size `pop_size` with `initial_result`
Args:
initial_result (GoalFunctionResult): Original text
pop_size (int): size of population
Returns:
population as `list[PopulationMember]`
"""
best_neighbors, prob_list = self._get_best_neighbors(
initial_result, initial_result
)
population = []
for _ in range(pop_size):
# Mutation step
random_result = np.random.choice(best_neighbors, 1, p=prob_list)[0]
population.append(
PopulationMember(random_result.attacked_text, random_result)
)
return population
def perform_search(self, initial_result):
self._search_over = False
population = self._initialize_population(initial_result, self.pop_size)
# Initialize up velocities of each word for each population
v_init = np.random.uniform(-self.v_max, self.v_max, self.pop_size)
velocities = np.array(
[
[v_init[t] for _ in range(initial_result.attacked_text.num_words)]
for t in range(self.pop_size)
]
)
global_elite = max(population, key=lambda x: x.score)
if (
self._search_over
or global_elite.result.goal_status == GoalFunctionResultStatus.SUCCEEDED
):
return global_elite.result
local_elites = copy.copy(population)
# start iterations
for i in range(self.max_iters):
omega = (self.omega_1 - self.omega_2) * (
self.max_iters - i
) / self.max_iters + self.omega_2
C1 = self.c1_origin - i / self.max_iters * (self.c1_origin - self.c2_origin)
C2 = self.c2_origin + i / self.max_iters * (self.c1_origin - self.c2_origin)
P1 = C1
P2 = C2
for k in range(len(population)):
# calculate the probability of turning each word
pop_mem_words = population[k].words
local_elite_words = local_elites[k].words
assert len(pop_mem_words) == len(
local_elite_words
), "PSO word length mismatch!"
for d in range(len(pop_mem_words)):
velocities[k][d] = omega * velocities[k][d] + (1 - omega) * (
self._equal(pop_mem_words[d], local_elite_words[d])
+ self._equal(pop_mem_words[d], global_elite.words[d])
)
turn_prob = utils.sigmoid(velocities[k])
if np.random.uniform() < P1:
# Move towards local elite
population[k] = self._turn(
local_elites[k],
population[k],
turn_prob,
initial_result.attacked_text,
)
if np.random.uniform() < P2:
# Move towards global elite
population[k] = self._turn(
global_elite,
population[k],
turn_prob,
initial_result.attacked_text,
)
# Check if there is any successful attack in the current population
pop_results, self._search_over = self.get_goal_results(
[p.attacked_text for p in population]
)
if self._search_over:
# if `get_goal_results` gets cut short by query budget, resize population
population = population[: len(pop_results)]
for k in range(len(pop_results)):
population[k].result = pop_results[k]
top_member = max(population, key=lambda x: x.score)
if (
self._search_over
or top_member.result.goal_status == GoalFunctionResultStatus.SUCCEEDED
):
return top_member.result
# Mutation based on the current change rate
for k in range(len(population)):
change_ratio = initial_result.attacked_text.words_diff_ratio(
population[k].attacked_text
)
# Referred from the original source code
p_change = 1 - 2 * change_ratio
if np.random.uniform() < p_change:
self._perturb(population[k], initial_result)
if self._search_over:
break
# Check if there is any successful attack in the current population
top_member = max(population, key=lambda x: x.score)
if (
self._search_over
or top_member.result.goal_status == GoalFunctionResultStatus.SUCCEEDED
):
return top_member.result
# Update the elite if the score is increased
for k in range(len(population)):
if population[k].score > local_elites[k].score:
local_elites[k] = copy.copy(population[k])
if top_member.score > global_elite.score:
global_elite = copy.copy(top_member)
return global_elite.result
def check_transformation_compatibility(self, transformation):
"""The genetic algorithm is specifically designed for word
substitutions."""
return transformation_consists_of_word_swaps(transformation)
@property
def is_black_box(self):
return True
def extra_repr_keys(self):
return ["pop_size", "max_iters", "post_turn_check", "max_turn_retries"]
def normalize(n):
n = np.array(n)
n[n < 0] = 0
s = np.sum(n)
if s == 0:
return np.ones(len(n)) / len(n)
else:
return n / s
|