InfoCensor

InfoCensor: An Information-Theoretic Framework against Sensitive Attribute Inference and Demographic Disparity

This codebase is written on top of pytorch (python3)

Some required packages are listed in requirement.txt

Datasets

German

The .csv file in german_data directory

Health Heritage

The .csv file in health_data directory

UTKFace

Download the UTKFace dataset from https://susanqq.github.io/UTKFace/ (Note that we use the utkface directory in the "Aligned and Cropped Faces") Replace the data path in "load_utkface.py". Note that for UTKFace and Twitter, it is better to use cuda to accelerate training (--use-cuda)

Twitter:

Get tokenizer from https://github.com/erikavaris/tokenizer pip install git+https://github.com/erikavaris/tokenizer.git You also need torchtext (find the version that matches your torch version)

Baselines without defense (code files are in the "defenses" directory)

python no_defense.py --dataset <dataset> --target-attr <the target attribute> --sensitive-attr <the sensitive attribute>

Examples (If you do not have GPU, then do not use the argument --use-cuda):

python no_defense.py --dataset german --target-attr credit --sensitive-attr gender

python no_defense.py --dataset health --target-attr charlson --sensitive-attr age

python no_defense.py --dataset utkface --num-epochs 100 --target-attr age --sensitive-attr race --use-cuda

Examples for InfoCensor (It is better to add --use-cuda if you have a gpu)

python infocensor.py --dataset health --target-attr charlson --sensitive-attr age --info-lambda 0.5 --info-beta 0.0 --fair-kappa 0.5

Attacks (code files are in the "attacks" directory)

python baseline.py --dataset <health or utkface or twitter> --defense-method <adv_censor or info_censor> --defense-model-name <the model name (record certain hyperparameters) for the specific defense> --sensitive-attr <the sensitive attribute>

python decensor.py --defense-method <adv_censor or info_censor> --dataset <health or utkface>  --defense-model-name <the model name (record certain hyperparameters) for the specific defense> --sensitive-attr <the sensitive attribute>

Examples (It is better to add --use-cuda if you have a gpu)

python baseline.py --dataset health --defense-method no_defense --defense-model-name target_charlson_sensitive_age_num_features_128 --sensitive-attr age

python baseline.py --dataset health --defense-method infocensor --defense-model-name target_charlson_sensitive_age_num_features_128_lambda_0.5_beta_0.0_kappa_0.5 --sensitive-attr age