Robustness evaluation via natural typos
A method for evaluating robustness of one or more target neural network models using natural typos. The method includes receiving one or more natural typo generation rules associated with a first task associated with a first input document type, receiving a first target neural network model, and receiving a first document and corresponding its ground truth labels. The method further includes generating one or more natural typos for the first document based on the one or more natural typo generation rules, and providing, to the first target neural network model, a test document generated based on the first document and the one or more natural typos as an input document to generate a first output. A robustness evaluation result of the first target neural network model is generated based on a comparison between the output and the ground truth labels.
1. A method for evaluating robustness of one or more target neural network models, comprising:
receiving one or more natural typo generation rules associated with a first task associated with a first input document type;
receiving a first target neural network model;
receiving a first document and corresponding ground truth labels;
generating one or more natural typos for the first document based on the one or more natural typo generation rules, using gradient information of components of the first document, wherein the generating the one or more natural typos include:
segmenting the first document to generate a plurality of sub-word components;
generating gradient information for the plurality of sub-word components; and
generating the one or more natural typos based on the gradient information;
providing, to the first target neural network model, a test document generated based on the first document and the one or more natural typos as an input document to generate a first output; and
generating a robustness evaluation result of the first target neural network model based on a comparison between the output and the ground truth labels.
2. The method of claim 1 , wherein the first input document type includes human-typed documents, and
wherein the one or more natural typo generation rules are configured to generate typos mimicking human keyboard typos.
3. The method of claim 1 , wherein the generating the one or more natural typos based on the gradient information includes:
determining a first sub-word component having one of a maximum gradient and a minimum gradient based on the gradient information;
determining a first word of the first document based on the first sub-word component; and
generating a first natural typo based on the first word.
4. The method of claim 3 , wherein the generating the one or more natural typos based on the gradient information includes:
determining a second sub-word component having the other of the maximum gradient and the minimum gradient based on the gradient information;
determining a second word of the first document based on the second sub-word component; and
generating a second natural typo based on the second word.
5. The method of claim 4 , wherein the generating the one or more natural typos include:
determining a third word of the first document randomly; and
generating a third natural typo based on the third word.
6. The method of claim 3 , wherein the robustness evaluation result of the first target neural network model includes:
a first model evaluation accuracy based on natural typos generates based on the maximum gradient information; and
a second model evaluation accuracy based on natural typos generates based on the minimum gradient information.
7. A non-transitory machine-readable medium comprising a plurality of machine-readable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform a method comprising:
receiving one or more natural typo generation rules associated with a first task associated with a first input document type;
receiving a first target neural network model;
receiving a first document and corresponding ground truth labels;
generating one or more natural typos for the first document based on the one or more natural typo generation rules, using gradient information of components of the first document, wherein the generating the one or more natural typos include:
segmenting the first document to generate a plurality of sub-word components;
generating gradient information for the plurality of sub-word components; and
generating the one or more natural typos based on the gradient information;
providing, to the first target neural network model, a test document generated based on the first document and the one or more natural typos as an input document to generate a first output; and
generating a robustness evaluation result of the first target neural network model based on a comparison between the output and the ground truth labels.
8. The non-transitory machine-readable medium of claim 7 , wherein the first input document type includes human-typed documents, and
wherein the one or more natural typo generation rules are configured to generate typos mimicking human keyboard typos.
9. The non-transitory machine-readable medium of claim 7 , wherein the generating the one or more natural typos based on the gradient information FAX (949) 202-3001 includes:
determining a first sub-word component having one of a maximum gradient and a minimum gradient based on the gradient information;
determining a first word of the first document based on the first sub-word component; and
generating a first natural typo based on the first word.
10. The non-transitory machine-readable medium of claim 9 , wherein the generating the one or more natural typos based on the gradient information includes:
determining a second sub-word component having the other of the maximum gradient and the minimum gradient based on the gradient information;
determining a second word of the first document based on the second sub-word component; and
generating a second natural typo based on the second word.
11. The non-transitory machine-readable medium of claim 10 , wherein the generating the one or more natural typos include:
determining a third word of the first document randomly; and
generating a third natural typo based on the third word.
12. The non-transitory machine-readable medium of claim 9 , wherein the robustness evaluation result of the first target neural network model includes:
a first model evaluation accuracy based on natural typos generates based on the maximum gradient information; and
a second model evaluation accuracy based on natural typos generates based on the minimum gradient information.
13. A system, comprising:
a non-transitory memory; and
one or more hardware processors coupled to the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform a method comprising:
receiving one or more natural typo generation rules associated with a first task associated with a first input document type;
receiving a first target neural network model;
receiving a first document and corresponding ground truth labels;
generating one or more natural typos for the first document based on the one or more natural typo generation rules, using gradient information of components of the first document, wherein the generating the one or more natural typos include:
segmenting the first document to generate a plurality of sub-word components;
generating gradient information for the plurality of sub-word components; and
generating the one or more natural typos based on the gradient information;
providing, to the first target neural network model, a test document generated based on the first document and the one or more natural typos as an input document to generate a first output; and
generating a robustness evaluation result of the first target neural network model based on a comparison between the output and the ground truth labels.
14. The system of claim 13 , wherein the first input document type includes human-typed documents, and
wherein the one or more natural typo generation rules are configured to generate typos mimicking human keyboard typos.
15. The system of claim 14 , wherein the generating the one or more natural typos based on the gradient information includes:
determining a first sub-word component having one of a maximum gradient and a minimum gradient based on the gradient information;
determining a first word of the first document based on the first sub-word component; and
generating a first natural typo based on the first word.
16. The system of claim 15 , wherein the generating the one or more natural typos based on the gradient information includes:
determining a second sub-word component having the other of the maximum gradient and the minimum gradient based on the gradient information;
determining a second word of the first document based on the second sub-word component; and
generating a second natural typo based on the second word.
17. The system of claim 15 , wherein the robustness evaluation result of the first target neural network model includes:
a first model evaluation accuracy based on natural typos generates based on the maximum gradient information; and
a second model evaluation accuracy based on natural typos generates based on the minimum gradient information.