Upload GreedyMultyGeneration.py
Browse files- GreedyMultyGeneration.py +721 -0
GreedyMultyGeneration.py
ADDED
@@ -0,0 +1,721 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import random
|
2 |
+
from textattack.search_methods import SearchMethod
|
3 |
+
from textattack.goal_function_results import GoalFunctionResultStatus
|
4 |
+
|
5 |
+
class GreedyMultipleGeneration(SearchMethod):
|
6 |
+
def __init__(
|
7 |
+
self,
|
8 |
+
wir_method="delete",
|
9 |
+
k=30,
|
10 |
+
embed=None,
|
11 |
+
file=None,
|
12 |
+
rollback_level=3,
|
13 |
+
naive=False,
|
14 |
+
clust=None,
|
15 |
+
train_file="train_file.csv",
|
16 |
+
):
|
17 |
+
self.wir_method = wir_method
|
18 |
+
self.k = k # maximum iterations
|
19 |
+
self.embed = embed # universal sentence encoder
|
20 |
+
self.file = file # similarity file to store the textual similarity
|
21 |
+
self.naive = naive
|
22 |
+
self.rollback_level = rollback_level
|
23 |
+
self.successful_attacks = {}
|
24 |
+
self.clust = clust
|
25 |
+
|
26 |
+
def _get_index_order(self, initial_text, indices_to_order):
|
27 |
+
"""Returns word indices of ``initial_text`` in descending order of
|
28 |
+
importance."""
|
29 |
+
|
30 |
+
if "unk" in self.wir_method:
|
31 |
+
leave_one_texts = [
|
32 |
+
initial_text.replace_word_at_index(i, "[UNK]") for i in indices_to_order
|
33 |
+
]
|
34 |
+
leave_one_results, search_over = self.get_goal_results(leave_one_texts)
|
35 |
+
index_scores = np.array([result.score for result in leave_one_results])
|
36 |
+
|
37 |
+
elif "delete" in self.wir_method:
|
38 |
+
leave_one_texts = [
|
39 |
+
initial_text.delete_word_at_index(i) for i in indices_to_order
|
40 |
+
]
|
41 |
+
leave_one_results, search_over = self.get_goal_results(leave_one_texts)
|
42 |
+
# print(f"leave_one_results : {leave_one_results}")
|
43 |
+
# print(f"search_over : {search_over}")
|
44 |
+
|
45 |
+
index_scores = np.array([result.score for result in leave_one_results])
|
46 |
+
|
47 |
+
elif "weighted-saliency" in self.wir_method:
|
48 |
+
# first, compute word saliency
|
49 |
+
leave_one_texts = [
|
50 |
+
initial_text.replace_word_at_index(i, "unk") for i in indices_to_order
|
51 |
+
]
|
52 |
+
leave_one_results, search_over = self.get_goal_results(leave_one_texts)
|
53 |
+
saliency_scores = np.array([result.score for result in leave_one_results])
|
54 |
+
|
55 |
+
softmax_saliency_scores = softmax(
|
56 |
+
torch.Tensor(saliency_scores), dim=0
|
57 |
+
).numpy()
|
58 |
+
|
59 |
+
# compute the largest change in score we can find by swapping each word
|
60 |
+
delta_ps = []
|
61 |
+
for idx in indices_to_order:
|
62 |
+
# Exit Loop when search_over is True - but we need to make sure delta_ps
|
63 |
+
# is the same size as softmax_saliency_scores
|
64 |
+
if search_over:
|
65 |
+
delta_ps = delta_ps + [0.0] * (
|
66 |
+
len(softmax_saliency_scores) - len(delta_ps)
|
67 |
+
)
|
68 |
+
break
|
69 |
+
|
70 |
+
transformed_text_candidates = self.get_transformations(
|
71 |
+
initial_text,
|
72 |
+
original_text=initial_text,
|
73 |
+
indices_to_modify=[idx],
|
74 |
+
)
|
75 |
+
if not transformed_text_candidates:
|
76 |
+
# no valid synonym substitutions for this word
|
77 |
+
delta_ps.append(0.0)
|
78 |
+
continue
|
79 |
+
swap_results, search_over = self.get_goal_results(
|
80 |
+
transformed_text_candidates
|
81 |
+
)
|
82 |
+
score_change = [result.score for result in swap_results]
|
83 |
+
if not score_change:
|
84 |
+
delta_ps.append(0.0)
|
85 |
+
continue
|
86 |
+
max_score_change = np.max(score_change)
|
87 |
+
delta_ps.append(max_score_change)
|
88 |
+
|
89 |
+
index_scores = softmax_saliency_scores * np.array(delta_ps)
|
90 |
+
|
91 |
+
elif "gradient" in self.wir_method:
|
92 |
+
victim_model = self.get_victim_model()
|
93 |
+
|
94 |
+
index_scores = np.zeros(len(indices_to_order))
|
95 |
+
grad_output = victim_model.get_grad(initial_text.tokenizer_input)
|
96 |
+
gradient = grad_output["gradient"]
|
97 |
+
word2token_mapping = initial_text.align_with_model_tokens(victim_model)
|
98 |
+
for i, index in enumerate(indices_to_order):
|
99 |
+
matched_tokens = word2token_mapping[index]
|
100 |
+
if not matched_tokens:
|
101 |
+
index_scores[i] = 0.0
|
102 |
+
else:
|
103 |
+
agg_grad = np.mean(gradient[matched_tokens], axis=0)
|
104 |
+
index_scores[i] = np.linalg.norm(agg_grad, ord=1)
|
105 |
+
|
106 |
+
search_over = False
|
107 |
+
|
108 |
+
index_order = np.array(indices_to_order)[(-index_scores).argsort()]
|
109 |
+
index_scores = sorted(index_scores, reverse=True)
|
110 |
+
return index_order, search_over, index_scores
|
111 |
+
|
112 |
+
# This present a rollback for reducing perturbation only
|
113 |
+
def swap_to_origin(self, cur_result, initial_result, index):
|
114 |
+
"""Replace the chosen word with it origin a return a result instance"""
|
115 |
+
new_attacked_text = cur_result.attacked_text.replace_word_at_index(
|
116 |
+
index, initial_result.attacked_text.words[index]
|
117 |
+
)
|
118 |
+
result, _ = self.get_goal_results([new_attacked_text])
|
119 |
+
return result[0]
|
120 |
+
|
121 |
+
def check_synonym_validity(
|
122 |
+
ind, ind_synonym, Synonym_indices, Current_attacked_Results, j, synonym
|
123 |
+
):
|
124 |
+
"""Checks if a synonym is valid for a given index in the attacked text.
|
125 |
+
|
126 |
+
Args:
|
127 |
+
ind: The index of the word in the attacked text.
|
128 |
+
ind_synonym: The index of the synonym in the list of synonyms.
|
129 |
+
Synonym_indices: A dictionary of synonym indices.
|
130 |
+
Current_attacked_Results: A list of AttackedResult objects.
|
131 |
+
j: The index of the current AttackedResult object in the list.
|
132 |
+
synonym: The synonym to check.
|
133 |
+
|
134 |
+
Returns:
|
135 |
+
True if the synonym is valid, False otherwise."""
|
136 |
+
|
137 |
+
# Check if the synonym has already been chosen.
|
138 |
+
if (ind, ind_synonym) in Synonym_indices:
|
139 |
+
return False
|
140 |
+
|
141 |
+
# Get the current attacked text and its words.
|
142 |
+
current_attacked_text = Current_attacked_Results[j].attacked_text
|
143 |
+
current_attacked_words = current_attacked_text.words
|
144 |
+
|
145 |
+
# Check if the synonym is already present in the attacked text.
|
146 |
+
if synonym in current_attacked_words[ind]:
|
147 |
+
return False
|
148 |
+
|
149 |
+
return True
|
150 |
+
|
151 |
+
def generate_naive_attack(self, initial_result):
|
152 |
+
curent_result = initial_result
|
153 |
+
# dict of preturbed indexes with theire scores on on the original text
|
154 |
+
perturbed_indexes = {}
|
155 |
+
# possible synonyms of each index with theire scores on the original text to reduce avg num queries
|
156 |
+
synonyms = {}
|
157 |
+
# to track indexes with no transformation so we avoid recalculate them to reduce avg num queries
|
158 |
+
non_usefull_indexes = []
|
159 |
+
attacked_text = initial_result.attacked_text
|
160 |
+
_, indices_to_order = self.get_indices_to_order(attacked_text)
|
161 |
+
|
162 |
+
# Sort words by order of importance
|
163 |
+
|
164 |
+
index_order, search_over, _ = self._get_index_order(
|
165 |
+
attacked_text, indices_to_order
|
166 |
+
)
|
167 |
+
|
168 |
+
# iterate through words by theire importance
|
169 |
+
for index in index_order:
|
170 |
+
if search_over:
|
171 |
+
break
|
172 |
+
transformed_text_candidates = self.get_transformations(
|
173 |
+
curent_result.attacked_text,
|
174 |
+
original_text=initial_result.attacked_text,
|
175 |
+
indices_to_modify=[index],
|
176 |
+
)
|
177 |
+
|
178 |
+
if len(transformed_text_candidates) == 0:
|
179 |
+
# track unusefull words to optimize the code .
|
180 |
+
non_usefull_indexes.append(index)
|
181 |
+
continue
|
182 |
+
else:
|
183 |
+
results, search_over = self.get_goal_results(
|
184 |
+
transformed_text_candidates
|
185 |
+
)
|
186 |
+
|
187 |
+
max_result = max(results, key=lambda x: x.score)
|
188 |
+
|
189 |
+
if max_result.score > curent_result.score:
|
190 |
+
if self.naive == False:
|
191 |
+
# store perturbed indexes with theire score
|
192 |
+
perturbed_indexes[index] = max_result.score - curent_result.score
|
193 |
+
# add all synonyms except the one we ve been using
|
194 |
+
synonyms[index] = [
|
195 |
+
(results[i].score, trans.words[index])
|
196 |
+
for i, trans in enumerate(transformed_text_candidates)
|
197 |
+
if trans.words[index] != max_result.attacked_text.words[index]
|
198 |
+
]
|
199 |
+
|
200 |
+
curent_result = max_result
|
201 |
+
|
202 |
+
if curent_result.goal_status == GoalFunctionResultStatus.SUCCEEDED:
|
203 |
+
return (
|
204 |
+
curent_result,
|
205 |
+
perturbed_indexes,
|
206 |
+
non_usefull_indexes,
|
207 |
+
synonyms,
|
208 |
+
curent_result.goal_status,
|
209 |
+
)
|
210 |
+
|
211 |
+
return (
|
212 |
+
curent_result,
|
213 |
+
perturbed_indexes,
|
214 |
+
non_usefull_indexes,
|
215 |
+
synonyms,
|
216 |
+
curent_result.goal_status,
|
217 |
+
)
|
218 |
+
|
219 |
+
# TODO we can add depth to track how many words rolled back for more statistics
|
220 |
+
|
221 |
+
def perturbed_index_swap(
|
222 |
+
self,
|
223 |
+
initial_result,
|
224 |
+
curent_result,
|
225 |
+
non_perturbed_indexes,
|
226 |
+
perturbed_indexes,
|
227 |
+
synonyms,
|
228 |
+
steps,
|
229 |
+
):
|
230 |
+
past_curent_result = curent_result
|
231 |
+
# the index with minimum perturbation
|
232 |
+
rollback_found = False
|
233 |
+
steps = min(steps, len(perturbed_indexes) - 1)
|
234 |
+
sucsefull_attacks = []
|
235 |
+
for _ in range(steps):
|
236 |
+
# TODO getting the least important perturbated word in the new attacked sample costs a lot
|
237 |
+
rollback_index = min(perturbed_indexes, key=perturbed_indexes.get)
|
238 |
+
# TODO remove from perturbed_indexes list and add it to non_perturbed_indexes but with punalitié
|
239 |
+
# how punalité should look like ? it could be at the end of the quee with visited flag
|
240 |
+
# or we can just eliminate it .
|
241 |
+
perturbed_indexes.pop(rollback_index, None)
|
242 |
+
for index in non_perturbed_indexes:
|
243 |
+
# early returning
|
244 |
+
if len(perturbed_indexes) == 1:
|
245 |
+
return (
|
246 |
+
curent_result,
|
247 |
+
non_perturbed_indexes,
|
248 |
+
perturbed_indexes,
|
249 |
+
synonyms,
|
250 |
+
sucsefull_attacks,
|
251 |
+
rollback_found,
|
252 |
+
)
|
253 |
+
|
254 |
+
# get candidates for non perturbed word
|
255 |
+
transformed_text_candidates = self.get_transformations(
|
256 |
+
curent_result.attacked_text,
|
257 |
+
original_text=initial_result.attacked_text,
|
258 |
+
indices_to_modify=[index],
|
259 |
+
)
|
260 |
+
|
261 |
+
if len(transformed_text_candidates) == 0:
|
262 |
+
non_perturbed_indexes.remove(index)
|
263 |
+
continue # wa7ed ma chaf wa7ed
|
264 |
+
|
265 |
+
results, _ = self.get_goal_results(transformed_text_candidates)
|
266 |
+
|
267 |
+
# we add one perturbed word
|
268 |
+
max_result = max(results, key=lambda x: x.score)
|
269 |
+
for res in results:
|
270 |
+
if res.score > curent_result.score:
|
271 |
+
if res.goal_status == GoalFunctionResultStatus.SUCCEEDED:
|
272 |
+
synonyms = self.update_synonyms(
|
273 |
+
synonyms=synonyms,
|
274 |
+
index_to_add=index,
|
275 |
+
index_to_remove=None,
|
276 |
+
curent_result=res,
|
277 |
+
results=results,
|
278 |
+
transformed_text_candidates=transformed_text_candidates,
|
279 |
+
)
|
280 |
+
# stock this sucssefull attack
|
281 |
+
sucsefull_attacks.append(res)
|
282 |
+
# we get better score
|
283 |
+
if max_result.score > curent_result.score:
|
284 |
+
# eplore minimum perturbation on the original text
|
285 |
+
inferior = min(perturbed_indexes, key=perturbed_indexes.get)
|
286 |
+
non_perturbed_indexes.remove(index) # remove perturbed index
|
287 |
+
|
288 |
+
perturbed_indexes[index] = max_result.score - curent_result.score
|
289 |
+
# restore one perturbed
|
290 |
+
result_rollback = self.swap_to_origin(
|
291 |
+
max_result, initial_result, rollback_index
|
292 |
+
)
|
293 |
+
|
294 |
+
perturbed_indexes.pop(inferior, None)
|
295 |
+
|
296 |
+
new_attacked_text = (
|
297 |
+
result_rollback.attacked_text.replace_word_at_index(
|
298 |
+
inferior,
|
299 |
+
initial_result.attacked_text.words[inferior],
|
300 |
+
)
|
301 |
+
)
|
302 |
+
|
303 |
+
result, _ = self.get_goal_results([new_attacked_text])
|
304 |
+
|
305 |
+
result_rollback = max(result, key=lambda x: x.score)
|
306 |
+
for res in result:
|
307 |
+
|
308 |
+
if res.goal_status == GoalFunctionResultStatus.SUCCEEDED:
|
309 |
+
synonyms = self.update_synonyms(
|
310 |
+
synonyms,
|
311 |
+
index,
|
312 |
+
inferior,
|
313 |
+
res,
|
314 |
+
results,
|
315 |
+
transformed_text_candidates,
|
316 |
+
)
|
317 |
+
# stock this sucssefull attack
|
318 |
+
sucsefull_attacks.append(res)
|
319 |
+
if (
|
320 |
+
result_rollback.goal_status
|
321 |
+
== GoalFunctionResultStatus.SUCCEEDED
|
322 |
+
):
|
323 |
+
rollback_found = True
|
324 |
+
synonyms = self.update_synonyms(
|
325 |
+
synonyms,
|
326 |
+
index,
|
327 |
+
inferior,
|
328 |
+
result_rollback,
|
329 |
+
results,
|
330 |
+
transformed_text_candidates,
|
331 |
+
)
|
332 |
+
curent_result = result_rollback
|
333 |
+
|
334 |
+
if rollback_found:
|
335 |
+
return (
|
336 |
+
curent_result,
|
337 |
+
non_perturbed_indexes,
|
338 |
+
perturbed_indexes,
|
339 |
+
synonyms,
|
340 |
+
sucsefull_attacks,
|
341 |
+
rollback_found,
|
342 |
+
)
|
343 |
+
return (
|
344 |
+
past_curent_result,
|
345 |
+
non_perturbed_indexes,
|
346 |
+
perturbed_indexes,
|
347 |
+
synonyms,
|
348 |
+
sucsefull_attacks,
|
349 |
+
rollback_found,
|
350 |
+
)
|
351 |
+
|
352 |
+
def update_synonyms(
|
353 |
+
self,
|
354 |
+
synonyms,
|
355 |
+
index_to_add=None,
|
356 |
+
index_to_remove=None,
|
357 |
+
curent_result=None,
|
358 |
+
results=None,
|
359 |
+
transformed_text_candidates=None,
|
360 |
+
):
|
361 |
+
"""Return an updated list of synonyms"""
|
362 |
+
if index_to_remove in synonyms and len(synonyms[index_to_remove]) != 0:
|
363 |
+
# remove the used synonym of certain index
|
364 |
+
synonyms[index_to_remove] = [
|
365 |
+
syn
|
366 |
+
for syn in synonyms[index_to_remove]
|
367 |
+
if syn[1] != curent_result.attacked_text.words[index_to_remove]
|
368 |
+
]
|
369 |
+
|
370 |
+
# add synonyms of new perturbated word with their score
|
371 |
+
if index_to_add is not None and transformed_text_candidates is not None:
|
372 |
+
synonyms[index_to_add] = [
|
373 |
+
(results[i].score, trans.words[index_to_add])
|
374 |
+
for i, trans in enumerate(transformed_text_candidates)
|
375 |
+
if trans.words[index_to_add]
|
376 |
+
!= curent_result.attacked_text.words[index_to_add]
|
377 |
+
]
|
378 |
+
|
379 |
+
return synonyms
|
380 |
+
|
381 |
+
def get_non_perturbed_indexes(
|
382 |
+
self, initial_result, perturbed_indexes, non_usefull_indexes
|
383 |
+
):
|
384 |
+
"""Return a list of non perturbed indexes"""
|
385 |
+
all_indexes = set(range(len(initial_result.attacked_text.words)))
|
386 |
+
perturbed_indexes_set = set(perturbed_indexes.keys())
|
387 |
+
non_usefull_indexes_set = set(non_usefull_indexes)
|
388 |
+
non_perturbed_indexes = list(
|
389 |
+
all_indexes - perturbed_indexes_set - non_usefull_indexes_set
|
390 |
+
)
|
391 |
+
return non_perturbed_indexes
|
392 |
+
|
393 |
+
def perform_search(self, initial_result):
|
394 |
+
|
395 |
+
(
|
396 |
+
curent_result,
|
397 |
+
perturbed_indexes,
|
398 |
+
non_usefull_indexes,
|
399 |
+
synonyms,
|
400 |
+
goal_statut,
|
401 |
+
) = self.generate_naive_attack(initial_result)
|
402 |
+
sucsefull_attacks = [curent_result]
|
403 |
+
|
404 |
+
new_curent_sucsefull_attacks = [curent_result]
|
405 |
+
if not self.naive:
|
406 |
+
# perturbed_index_swap is our 1s priority (in case of attack succeed goal_statut = 0 )
|
407 |
+
for i in range(self.k):
|
408 |
+
non_perturbed_indexes = self.get_non_perturbed_indexes(
|
409 |
+
initial_result, perturbed_indexes, non_usefull_indexes
|
410 |
+
)
|
411 |
+
if len(new_curent_sucsefull_attacks) != 0:
|
412 |
+
# how to decide on the next text to be treated here we work on the the one with max score
|
413 |
+
curent_result = max(
|
414 |
+
new_curent_sucsefull_attacks, key=lambda x: x.score
|
415 |
+
)
|
416 |
+
new_curent_sucsefull_attacks.remove(curent_result)
|
417 |
+
else:
|
418 |
+
curent_result, synonyms, synonym_found = self.swap_to_synonym(
|
419 |
+
curent_result, synonyms, perturbed_indexes
|
420 |
+
)
|
421 |
+
if synonym_found == True:
|
422 |
+
sucsefull_attacks.append(curent_result)
|
423 |
+
new_curent_sucsefull_attacks.append(curent_result)
|
424 |
+
continue
|
425 |
+
|
426 |
+
else:
|
427 |
+
non_perturbed_indexes = self.get_non_perturbed_indexes(
|
428 |
+
initial_result, perturbed_indexes, non_usefull_indexes
|
429 |
+
)
|
430 |
+
(
|
431 |
+
non_perturbed_indexes,
|
432 |
+
perturbed_indexes,
|
433 |
+
synonyms,
|
434 |
+
max_result,
|
435 |
+
sample_found,
|
436 |
+
) = self.random_selection(
|
437 |
+
non_perturbed_indexes,
|
438 |
+
perturbed_indexes,
|
439 |
+
synonyms,
|
440 |
+
curent_result,
|
441 |
+
initial_result,
|
442 |
+
)
|
443 |
+
|
444 |
+
if sample_found == True:
|
445 |
+
new_curent_sucsefull_attacks.append(max_result)
|
446 |
+
sucsefull_attacks.append(curent_result)
|
447 |
+
|
448 |
+
else:
|
449 |
+
break
|
450 |
+
if i % 3 == 0:
|
451 |
+
non_perturbed_indexes = self.get_non_perturbed_indexes(
|
452 |
+
initial_result, perturbed_indexes, non_usefull_indexes
|
453 |
+
)
|
454 |
+
(
|
455 |
+
non_perturbed_indexes,
|
456 |
+
perturbed_indexes,
|
457 |
+
synonyms,
|
458 |
+
max_result,
|
459 |
+
sample_found,
|
460 |
+
) = self.random_selection(
|
461 |
+
non_perturbed_indexes,
|
462 |
+
perturbed_indexes,
|
463 |
+
synonyms,
|
464 |
+
curent_result,
|
465 |
+
initial_result,
|
466 |
+
)
|
467 |
+
if sample_found == True:
|
468 |
+
new_curent_sucsefull_attacks.append(max_result)
|
469 |
+
sucsefull_attacks.append(curent_result)
|
470 |
+
|
471 |
+
if len(perturbed_indexes) > 1 and not goal_statut:
|
472 |
+
non_perturbed_indexes = self.get_non_perturbed_indexes(
|
473 |
+
initial_result, perturbed_indexes, non_usefull_indexes
|
474 |
+
)
|
475 |
+
(
|
476 |
+
curent_result,
|
477 |
+
non_perturbed_indexes,
|
478 |
+
perturbed_indexes,
|
479 |
+
synonyms,
|
480 |
+
sucsefull_attacks_partial,
|
481 |
+
rollback_found,
|
482 |
+
) = self.perturbed_index_swap(
|
483 |
+
initial_result,
|
484 |
+
curent_result,
|
485 |
+
non_perturbed_indexes,
|
486 |
+
perturbed_indexes,
|
487 |
+
synonyms,
|
488 |
+
steps=self.rollback_level,
|
489 |
+
)
|
490 |
+
if len(sucsefull_attacks_partial) != 0:
|
491 |
+
sucsefull_attacks.extend(sucsefull_attacks_partial)
|
492 |
+
new_curent_sucsefull_attacks.extend(sucsefull_attacks_partial)
|
493 |
+
# Action 2: the case where no rollback found we try to swap synonym and we aim to get better result
|
494 |
+
if rollback_found == False:
|
495 |
+
curent_result, synonyms, synonym_found = self.swap_to_synonym(
|
496 |
+
curent_result, synonyms, perturbed_indexes
|
497 |
+
)
|
498 |
+
if synonym_found == True:
|
499 |
+
sucsefull_attacks.append(curent_result)
|
500 |
+
new_curent_sucsefull_attacks.append(curent_result)
|
501 |
+
|
502 |
+
# if it's a failed attack we give chance for an other synonym
|
503 |
+
# we will pass it for now because no improvment were found
|
504 |
+
"""elif goal_statut == 1:
|
505 |
+
curent_result, synonyms, goal_statut = self.swap_to_synonym(
|
506 |
+
curent_result, synonyms, perturbed_indexes
|
507 |
+
)"""
|
508 |
+
|
509 |
+
if goal_statut == 0:
|
510 |
+
sucsefull_attacks_text_scores = []
|
511 |
+
sucsefull_attacks_text_scores = [
|
512 |
+
(atk.attacked_text, atk.score)
|
513 |
+
for atk in sucsefull_attacks
|
514 |
+
if atk.score > 0.5
|
515 |
+
]
|
516 |
+
sucsefull_attacks_text_scores = list(set(sucsefull_attacks_text_scores))
|
517 |
+
|
518 |
+
self.successful_attacks[initial_result.attacked_text] = (
|
519 |
+
sucsefull_attacks_text_scores
|
520 |
+
)
|
521 |
+
ground_truth_output = sucsefull_attacks[0].ground_truth_output
|
522 |
+
|
523 |
+
self.save_to_train(
|
524 |
+
self,
|
525 |
+
initial_result.attacked_text,
|
526 |
+
sucsefull_attacks_text_scores,
|
527 |
+
ground_truth_output,
|
528 |
+
)
|
529 |
+
|
530 |
+
try:
|
531 |
+
best_result = self.min_perturbation(
|
532 |
+
sucsefull_attacks, initial_result.attacked_text
|
533 |
+
)
|
534 |
+
return best_result
|
535 |
+
except:
|
536 |
+
return curent_result
|
537 |
+
|
538 |
+
def save_to_train(
|
539 |
+
self,
|
540 |
+
original_text,
|
541 |
+
sucsefull_attacks_text_scores,
|
542 |
+
ground_truth_output,
|
543 |
+
train_file,
|
544 |
+
):
|
545 |
+
successful_attacks = {
|
546 |
+
original_text.attacked_text: sucsefull_attacks_text_scores
|
547 |
+
}
|
548 |
+
self.save_to_JSON(filename="temp.json", successful_attacks=successful_attacks)
|
549 |
+
|
550 |
+
self.pipeline(ground_truth_output, train_file)
|
551 |
+
|
552 |
+
def pipeline(self, ground_truth_output, train_file):
|
553 |
+
clust = self.clust
|
554 |
+
clust.file_ = "temp.json"
|
555 |
+
sentence_embedding_vectors, masks, scores = clust.prepare_sentences()
|
556 |
+
|
557 |
+
unified_mask = clust.get_global_unified_masks(masks=masks)
|
558 |
+
|
559 |
+
sentences = clust.apply_mask_on_global_vectors(
|
560 |
+
global_sentences=sentence_embedding_vectors, unified_masks=unified_mask
|
561 |
+
)
|
562 |
+
|
563 |
+
sentences = clust.global_matrix_to_global_sentences(
|
564 |
+
global_matrix_sentences=sentences
|
565 |
+
)
|
566 |
+
|
567 |
+
global_clustering = clust.find_global_best_clustering(
|
568 |
+
sentences, 10, "thumb-rule"
|
569 |
+
)
|
570 |
+
|
571 |
+
selected_samples = clust.global_select_diverce_sample(
|
572 |
+
scores, sentences, global_clustering
|
573 |
+
)
|
574 |
+
|
575 |
+
clust.save_csv(selected_samples, ground_truth_output, train_file)
|
576 |
+
|
577 |
+
def save_to_JSON(self, filename, successful_attacks):
|
578 |
+
data_list = []
|
579 |
+
input_dict = {}
|
580 |
+
for atk in successful_attacks:
|
581 |
+
successful_attacks_with_scores = [
|
582 |
+
(atk, score) for atk, score in successful_attacks[atk]
|
583 |
+
]
|
584 |
+
input_dict[" ".join(atk.words)] = successful_attacks_with_scores
|
585 |
+
for original, samples in input_dict.items():
|
586 |
+
samples_list = [
|
587 |
+
{"attacked_text": " ".join(text.words), "score": score}
|
588 |
+
for text, score in samples
|
589 |
+
]
|
590 |
+
data_list.append({"original": original, "samples": samples_list})
|
591 |
+
|
592 |
+
# Save the formatted data to a JSON file
|
593 |
+
with open(filename, "w") as json_file:
|
594 |
+
json.dump({"data": data_list}, json_file, indent=4)
|
595 |
+
|
596 |
+
def swap_to_synonym(self, curent_result, synonyms, perturbed_indexes):
|
597 |
+
# giving chance to the second synonym of the most perturbated word if exists !
|
598 |
+
found = False
|
599 |
+
for index in perturbed_indexes:
|
600 |
+
if index in synonyms and len(synonyms[index]) != 0:
|
601 |
+
# what about other indexes we may give them chance too !
|
602 |
+
# response : experiments shows that there is no much improvment taking in consideration the high increase of avg Q-num
|
603 |
+
synonym = max(synonyms[index], key=lambda x: x[0])
|
604 |
+
if synonym[0] > 0.8:
|
605 |
+
new_attacked_text = (
|
606 |
+
curent_result.attacked_text.replace_word_at_index(
|
607 |
+
index,
|
608 |
+
synonym[1],
|
609 |
+
)
|
610 |
+
)
|
611 |
+
curent_result.attacked_text = (
|
612 |
+
curent_result.attacked_text.replace_word_at_index(
|
613 |
+
index,
|
614 |
+
synonym[1],
|
615 |
+
)
|
616 |
+
)
|
617 |
+
|
618 |
+
synonyms = self.update_synonyms(
|
619 |
+
synonyms=synonyms,
|
620 |
+
index_to_remove=index,
|
621 |
+
curent_result=curent_result,
|
622 |
+
)
|
623 |
+
found = True
|
624 |
+
return curent_result, synonyms, found
|
625 |
+
|
626 |
+
# remove index with 0 synonymswithin the list
|
627 |
+
synonyms.pop(index, None)
|
628 |
+
|
629 |
+
return curent_result, synonyms, found
|
630 |
+
|
631 |
+
def min_perturbation(self, results, original_text):
|
632 |
+
# Initialize minimum score and result
|
633 |
+
min_score = float("inf")
|
634 |
+
min_result = None
|
635 |
+
original_text_splited = original_text.words
|
636 |
+
for result in results:
|
637 |
+
# Calculate perturbation as the number of words changed
|
638 |
+
attacked_text = result.attacked_text
|
639 |
+
perturbation = sum(
|
640 |
+
i != j for i, j in zip(original_text_splited, attacked_text.words)
|
641 |
+
)
|
642 |
+
|
643 |
+
# Update minimum score and result if necessary
|
644 |
+
if perturbation < min_score:
|
645 |
+
min_score = perturbation
|
646 |
+
min_result = result
|
647 |
+
|
648 |
+
return min_result
|
649 |
+
|
650 |
+
def check_transformation_compatibility(self, transformation):
|
651 |
+
"""Since it ranks words by their importance, the algorithm is
|
652 |
+
limited to word swap and deletion transformations."""
|
653 |
+
return transformation_consists_of_word_swaps_and_deletions(transformation)
|
654 |
+
|
655 |
+
def random_selection(
|
656 |
+
self,
|
657 |
+
non_perturbed_indexes,
|
658 |
+
perturbed_indexes,
|
659 |
+
synonyms,
|
660 |
+
curent_result,
|
661 |
+
initial_result,
|
662 |
+
):
|
663 |
+
max_iterations = len(non_perturbed_indexes)
|
664 |
+
sample_found = False
|
665 |
+
for _ in range(max_iterations):
|
666 |
+
random_index = random.choice(non_perturbed_indexes)
|
667 |
+
transformed_text_candidates = self.get_transformations(
|
668 |
+
curent_result.attacked_text,
|
669 |
+
original_text=initial_result.attacked_text,
|
670 |
+
indices_to_modify=[random_index],
|
671 |
+
)
|
672 |
+
if len(transformed_text_candidates) == 0:
|
673 |
+
non_perturbed_indexes.remove(random_index)
|
674 |
+
continue
|
675 |
+
|
676 |
+
results, _ = self.get_goal_results([transformed_text_candidates[0]])
|
677 |
+
|
678 |
+
# we add one perturbed word
|
679 |
+
max_result = max(results, key=lambda x: x.score)
|
680 |
+
sample_found = True
|
681 |
+
# update synonym
|
682 |
+
synonyms = self.update_synonyms(
|
683 |
+
synonyms=synonyms,
|
684 |
+
index_to_add=random_index,
|
685 |
+
curent_result=curent_result,
|
686 |
+
results=results,
|
687 |
+
transformed_text_candidates=[transformed_text_candidates[0]],
|
688 |
+
)
|
689 |
+
|
690 |
+
# penalty on existing indexes
|
691 |
+
for index in perturbed_indexes:
|
692 |
+
perturbed_indexes[index] = perturbed_indexes[index] * 0.9
|
693 |
+
|
694 |
+
perturbed_indexes[random_index] = max_result.score - curent_result.score
|
695 |
+
non_perturbed_indexes.remove(random_index)
|
696 |
+
|
697 |
+
return (
|
698 |
+
non_perturbed_indexes,
|
699 |
+
perturbed_indexes,
|
700 |
+
synonyms,
|
701 |
+
max_result,
|
702 |
+
sample_found,
|
703 |
+
)
|
704 |
+
|
705 |
+
return (
|
706 |
+
non_perturbed_indexes,
|
707 |
+
perturbed_indexes,
|
708 |
+
synonyms,
|
709 |
+
curent_result,
|
710 |
+
sample_found,
|
711 |
+
)
|
712 |
+
|
713 |
+
@property
|
714 |
+
def is_black_box(self):
|
715 |
+
if "gradient" in self.wir_method:
|
716 |
+
return False
|
717 |
+
else:
|
718 |
+
return True
|
719 |
+
|
720 |
+
def extra_repr_keys(self):
|
721 |
+
return ["wir_method"]
|