Skip to content

Robustness Toolbox

(aiverify.stock.robustness-toolbox) [source]

Description

This plugin generates a perturbed dataset using boundary attack algorithm on the test dataset.

Boundary Attack is an attack that starts by adding a large amount of noise to a data point intentionally to cause a model it misclassified by the model. This plugin uses Salt-and-pepper noise to create the large amount of noise. Then, it will reduce the amount of noise added while maintaining misclassification. This algorithm does not depend on the underlying model's architecture or parameters.

This algorithm is developed for image dataset but can also be used to create noise on tabular dataset. However, it is to note that testing on tabular dataset may warrant caution when interpreting the results as this is not well-tested.

Plugin Content

  • Algorithms
Name Description
Robustness Toolbox This algorithm generates a perturbed dataset using boundary attack algorithm on the test dataset
  • Widgets
Name Description
Description (Summary) To provide introduction, interpretation and recommendations for robustness testing
Bar Chart (Accuracy) To generate and display a bar chart of the orignal and perturbed dataset with interpretation of the results
Description (Technical) To provide introduction, bar chart, interpretation and recommendations for robustness testing with technical details

Using the Plugin in AI Verify

Data Preparation

This plugin was mainly designed for image datasets, but can also be used on tabular datasets.

For images:

For tabular:

Algorithm User Input(s)

Input Field Description Type
Annotated ground truth path For image datasets: An uploaded dataset containing image file names and the corresponding ground truth label
For tabular datasets: Select the ground truth dataset		 string
Name of column containing image file name For image datasets: Key in the name of the column containing the file names in the annotated ground truth dataset
For tabular datasets: Key in NA		 string

Sample use of the widgets

Robustness Toolbox Sample

More details

Algorithm input schema 
{
    "title": "Algorithm Plugin Input Arguments",
    "description": "A schema for algorithm plugin input arguments",
    "type": "object",
    "required": [
    ],
    "properties": {
        "annotated_ground_truth_path": {
            "title": "Annotated ground truth path",
            "description": "Annotated ground truth path",
            "type": "string",
            "ui:widget": "selectDataset"
        },
        "file_name_label": {
            "title": "Name of column containing image file names",
            "description": "Key in the name of the column containing the file names in the annotated ground truth dataset",
            "type": "string"
        }
    }
}
Algorithm output schema 
{
    "title": "Algorithm Plugin Output Arguments",
    "description": "A schema for algorithm plugin output arguments",
    "type": "object",
    "required": ["results"],
    "minProperties": 1,
    "properties": {
        "results": {
            "description": "Algorithm Output",
            "type": "object",
            "required": ["num_of_perturbed_samples", "org_performance", "perturbed_performance", "num_of_failed_perturbed_samples"],
            "properties": {
                "num_of_perturbed_samples": {
                    "description": "Number of final perturbed samples",
                    "type": "number"
                },
                "original": {
                    "description": "Performance for Original Dataset",
                    "type": "number"
                },
                "adversarial": {
                    "description": "Performance for Perturbed Dataset ",
                    "type": "number"
                },
                "num_of_failed_perturbed_samples": {
                    "description": "Number of samples that failed to generate perturbed samples",
                    "type": "number"
                }
            }
        }
    }
}