AEGIS - A Correlation-Based Data Masking Advisor for Data-Sharing Ecosystems

Masking Configuration Generation

To generate masking configurations for a dataset, use:

make generate_masking_configurations ORIGINAL_DATA="path/to/your/dataset.csv" CLASS_LABEL="target_column_name" TOTAL_CONFIGURATIONS=10 CONFIG_DIR="configs"

Parameters

• ORIGINAL_DATA - Path to the original CSV dataset
• CLASS_LABEL - Name of the target variable (excluded from masking)
• TOTAL_CONFIGURATIONS - Number of masking configurations to generate
• CONFIG_DIR - Directory where configurations will be stored (default: configs)

Output

The command will generate YAML configuration files in the configs directory. Each configuration includes:

1. Dataset path information
2. Masking functions to be applied to each attribute

Example configuration:

dataset:
  original_path: path/to/your/dataset.csv
  target_variable: target_column_name
masking:
  attributes:
    age:
    - function: generalize
      params:
        M: 3
    income:
    - function: suppress
      params: {}

Applying Data Masking

To apply a masking configuration and generate a masked dataset, use:

make mask_data MASKING_CONFIG="configs/config_1.yaml" MASKED_DATA="output/masked_dataset.csv"

Parameters

• MASKING_CONFIG - Path to a YAML masking configuration file
• MASKED_DATA - Path to the output masked dataset (including directory)

Output

The command will apply the specified masking configuration to the original dataset (defined in the configuration file) and save the masked data to the specified output location.

Computing Distributions

Before performing data reconstruction, you need to compute distributions. Use:

make compute_distribution DATASET="path/to/dataset.csv" MARGINALS="path/to/marginals.pkl" JOINT_DISTRIBUTION="path/to/joint_distribution.pkl" CLASS_LABEL="target_column_name"

Parameters

• DATASET - Path to the input dataset CSV file
• MARGINALS - Path to save marginals (frequency distributions for each attribute)
• JOINT_DISTRIBUTION - Path to save joint distributions with class label
• CLASS_LABEL - Name of the target/class variable column

Output

The command computes:

1. Marginals: frequency counts for each value in each attribute
2. Joint distributions: joint distributions between each attribute and the class label

These distributions are saved as pickle files and are required for the reconstruction algorithms.

Evaluating Predictive Utility Deviation (PUD)

To evaluate the PUD between masked and reconstructed distributions, use:

make evaluate_pud METRIC="chi2" MASKED_JOINT_DIST="path/to/masked.pkl" RECONSTRUCTED_JOINT_DIST="path/to/reconstructed.pkl"

Parameters

• METRIC - Metric to use for evaluation (options: chi2, mi, tvd, g3)
• MASKED_JOINT_DIST - Path to masked joint distribution pickle file
• RECONSTRUCTED_JOINT_DIST - Path to reconstructed joint distribution pickle file

Output

The command calculates and displays PUD across all attributes

Available metrics:

• chi2 - Chi-square statistic
• mi - Mutual Information
• tvd - Total Variation Distance
• g3 - G3 metric

Evaluating Model Performance

To run a machine learning model on a dataset and evaluate its performance, use:

make model DATA="path/to/dataset.csv" MODEL="path/to/model_file.py" TARGET="target_column_name"

Parameters

• DATA - Path to the dataset CSV file to evaluate
• MODEL - Path to the model Python file to use for evaluation
• TARGET - Name of the target variable column

Output

The command will train the specified model on the dataset and output its accuracy.