US 12,573,382 B2 · Adversarial language imitation with constrained exemplars

Abstract

Generally discussed herein are devices, systems, and methods for generating a phrase that is confusing to a language classifier. A method can include determining, by the LC, a first classification score (CS) of a prompt indicating whether the prompt is a first class or a second class, predicting, based on the prompt and by a pre-trained language model (PLM), likely next words and a corresponding probability for each of the likely next words, determining, by the LC, a second CS for each of the likely next words, determining, by an adversarial classifier, respective scores for each of the likely next words, the respective scores determined based on the first CS of the prompt, the second CS of the likely next words, and the probabilities of the likely next words, and selecting, by an adversarial classifier, a next word of the likely next words based on the respective scores.

Claims20

Dependent claims

A method for generating a phrase that is confusing for a language classifier (LC), the method comprising:

receiving from the LC, a first classification score (CS) of a prompt;
iteratively performing operations to generate a confusing prompt used for training another language classifier model, the operation comprising:
receiving, by a classifier, input including:
the first CS,
from a language model (LM) and based on a next word of the prompt, predicted likely next words for the prompt, and corresponding probabilities for the predicted likely next words,
from the LC and based on the predicted likely next words, a second CS of the predicted likely next words, the first and second CS indicate whether the prompt and the predicted likely next words are a first class or a second class, respectively;
determining, by the classifier and based on the input, respective confusion scores for the predicted likely next words;
selecting, by the classifier, a likely next word of the predicted likely next words associated with a maximum confusion score of the respective confusion scores;
and
appending the selected likely next word to the confusing prompt.

7 more elements

A compute system comprising:

a memory;
processing circuitry coupled to the memory, the processing circuitry configured to:
receiving from a language classifier (LC), a first classification score (CS) of a prompt;
iteratively performing operations to generate a confusing prompt used for training another language classifier model, the operations comprising:
receiving, by a classifier, input including the first CS and:
from a language model (LM) and based on a next word of the prompt, predicted likely next words for the prompt, and corresponding probabilities for the predicted likely next words,
from the LC and based on the predicted likely next words, a second CS of the predicted likely next words, the first and second CS indicate whether the prompt and the predicted likely next words are a first class or a second class, respectively;
determining, by the classifier and based on the input, respective confusion scores for the predicted likely next words:
selecting, by the classifier, a likely next word of the predicted likely next words associated with a maximum confusion score of the respective confusion scores;
and
appending the selected likely next word to the confusing prompt.

8 more elements

A machine-readable medium including instructions that, when executed by a machine, cause the machine to perform operations comprising:

receiving from a language classifier (LC), a first classification score (CS) of a prompt;
iteratively performing operations to generate a confusing prompt used for training another language classifier model, the operations comprising:
receiving, by a classifier, input including the first CS and:
from a language model (LM) and based on a next word of the prompt, predicted likely next words for the prompt, and corresponding probabilities for the predicted likely next words,
from the LC and based on the predicted likely next words, a second CS of the predicted likely next words, the first and second CS indicate whether the prompt and the predicted likely next words are a first class or a second class, respectively;
determining, by the classifier and based on the input, respective confusion scores for the predicted likely next words;
selecting, by the classifier, a likely next word of the predicted likely next words associated with a maximum confusion score of the respective confusion scores;
and
appending the selected likely next word to the confusing prompt.

6 more elements

Dependencies are inferred from the claim text.

Adversarial language imitation with constrained exemplars

Life of this patent

Adversarial language imitation with constrained exemplars

Life of this patent

Abstract

Claims20