Substitutes API¶

Abstract Classes¶

class WordSubstitute[source]¶

__call__(word, pos=None)[source]¶

In WordSubstitute, we return a list of words that are semantically similar to the input word.

Parameters

word (str) – A single word.
pos (Optional[str]) – POS tag of input word. Must be one of the following: ["adv", "adj", "noun", "verb", "other", None]

Returns

A list of words and their distance to original word (distance is a number between 0 and 1, with smaller indicating more similarity)

Raises

WordNotInDictionaryException – input word not in the dictionary of substitute algorithm
UnknownPOSException – invalid pos tagging

Return type

List[Tuple[str, float]]

class CharSubstitute[source]¶

__call__(char)[source]¶

Char-level substitute algorithm.

In CharSubstitute, we return a list of chars that are visually similar to the original word.

Parameters: char (str) – A signle char
Returns: A list of chars and distance to original char (distance is a number between 0 and 1, with smaller indicating more similarity).
Return type: List[Tuple[str, float]]

EmbedBasedSubstitute¶

class EmbedBasedSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(word2id, embedding, cosine=False, k=50, threshold=0.5, device=None)[source]¶

Embedding based word substitute.

Parameters

word2id (Dict[str, int]) – A dict maps words to indexes.
embedding (torch.Tensor) – A word embedding matrix.
cosine – If true then the cosine distance is used, otherwise the Euclidian distance is used.
threshold – Distance threshold. Default: 0.5
k – Top-k results to return. If k is None, all results will be returned. Default: 50
device – A pytocrh device for computing distances. Default: “cpu”

ChineseHowNetSubstitute¶

class ChineseHowNetSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(k=None)[source]¶

Chinese Sememe-based word substitute based on OpenHowNet. [pdf]

Parameters: k (Optional[int]) – Top-k results to return. If k is None, all results will be returned.
Package Requirements: OpenHowNet
Data Requirements: AttackAssist.HowNet
Language: chinese

ChineseWord2VecSubstitute¶

class ChineseWord2VecSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(cosine=False, threshold=0.5, k=50, device=None)[source]¶

Chinese word substitute based on word2vec.

Parameters

cosine (bool) – If true then the cosine distance is used, otherwise the Euclidian distance is used.
threshold (float) – Distance threshold. Default: 0.5
k (int) – Top-k results to return. If k is None, all results will be returned. Default: 50
device (Optional[Union[str, torch.device]]) – A pytocrh device for computing distances. Default: “cpu”

Data Requirements

AttackAssist.ChineseWord2Vec

Language

chinese

ChineseWordNetSubstitute¶

class ChineseWordNetSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(k=None)[source]¶

Chinese word substitute based on wordnet.

Parameters: k – Top-k results to return. If k is None, all results will be returned. Default: 50
Language: chinese

CounterFittedSubstitute¶

class CounterFittedSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(cosine=False, k=50, threshold=0.5, device=None)[source]¶

English word substitute based on Counter-fitting word vectors. [pdf]

Parameters

cosine (bool) – If true then the cosine distance is used, otherwise the Euclidian distance is used.
threshold (float) – Distance threshold. Default: 0.5
k (int) – Top-k results to return. If k is None, all results will be returned. Default: 50
device – A pytocrh device for computing distances. Default: “cpu”

Data Requirements

AttackAssist.CounterFit

Language

english

ChineseCiLinSubstitute¶

class ChineseCiLinSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(k=None)[source]¶

Chinese Sememe-based word substitute based CiLin.

Parameters: k (Optional[int]) – Top-k results to return. If k is None, all results will be returned.
Data Requirements: AttackAssist.CiLin
Language: chinese

GloveSubstitute¶

class GloveSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(cosine=False, k=50, threshold=0.5, device=None)[source]¶

English word substitute based on GloVe word vectors. [pdf]

Parameters

cosine – If true then the cosine distance is used, otherwise the Euclidian distance is used.
threshold – Distance threshold. Default: 0.5
k – Top-k results to return. If k is None, all results will be returned. Default: 50
device – A pytocrh device for computing distances. Default: “cpu”

Data Requirements

AttackAssist.GloVe

Language

english

HowNetSubstitute¶

class HowNetSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(k=None)[source]¶

English Sememe-based word substitute based on OpenHowNet. [pdf]

Parameters: k – Top-k results to return. If k is None, all results will be returned.
Data Requirements: AttackAssist.HownetSubstituteDict
Language: english

Word2VecSubstitute¶

class Word2VecSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(cosine=False, k=50, threshold=0.5, device=None)[source]¶

English word substitute based on word2vec.

Parameters

cosine – If true then the cosine distance is used, otherwise the Euclidian distance is used.
threshold – Distance threshold. Default: 0.5
k – Top-k results to return. If k is None, all results will be returned. Default: 50
device – A pytocrh device for computing distances. Default: “cpu”

Data Requirements

AttackAssist.GloVe

Language

english

WordNetSubstitute¶

class WordNetSubstitute(OpenAttack.attack_assist.substitute.word.WordSubstitute)[source]¶

__init__(k=None)[source]¶

English word substitute based on wordnet.

Parameters: k – Top-k results to return. If k is None, all results will be returned. Default: 50
Data Requirements: TProcess.NLTKWordNet
Language: english

ChineseFYHCharSubstitute¶

class ChineseFYHCharSubstitute(OpenAttack.attack_assist.substitute.char.CharSubstitute)[source]¶

__init__(k=None)[source]¶

Returns traditional, variant and Martian characters of the input character.

Parameters: k (Optional[int]) – Top-k results to return. If k is None, all results will be returned.
Data Requirements: AttackAssist.FYH
Language: chinese

ChineseSimCharSubstitute¶

class ChineseSimCharSubstitute(OpenAttack.attack_assist.substitute.char.CharSubstitute)[source]¶

__init__(k=None)[source]¶

Returns the chars that is visually similar to the input.

Parameters: k (Optional[int]) – Top-k results to return. If k is None, all results will be returned.
Data Requirements: AttackAssist.SIM
Language: chinese

DCESSubstitute¶

class DCESSubstitute(OpenAttack.attack_assist.substitute.char.CharSubstitute)[source]¶

__init__(k=12)[source]¶

Returns the chars that is visually similar to the input.

DCES substitute used in VIPERAttacker.

Parameters

k (int) – Top-k results to return. Default: k = 12

Data Requirements

AttackAssist.SIM

Language

english

Package Requirements

sklearn

ECESSubstitute¶

class ECESSubstitute(OpenAttack.attack_assist.substitute.char.CharSubstitute)[source]¶

__init__()[source]¶

Returns the chars that is visually similar to the input.

DCES substitute used in VIPERAttacker.

Data Requirements: AttackAssist.SIM
Language: english