label-protection

Code Repo for paper Label Leakage and Protection in Two-party Split Learning (ICLR 2022).

Requirements

• Python 3
• Tensorflow >2.0
• scikit-learn

Dataset

• For demonstration, we have provided a small portion (0.5%) of Criteo dataset in dataset/criteo.
• Interested readers can use above dataset to test our code.
• For the actual datasets (Avazu, ISIC), readers should download the datasets Criteo: https://www.kaggle.com/c/criteo-display-ad-challenge/data Avazu: https://www.kaggle.com/c/avazu-ctr-prediction/data ISIC: https://www.kaggle.com/c/siim-isic-melanoma-classification/data and use preprocess_criteo_subset.py preprocess_avazu.py and preprocess_ISIC.ipynb to preprocess the corresponding datasets.

Run

We provide a script for each dataset to test our protection methods.

• run_script_criteo.py for Criteo
• run_script_avazu.py for Avazu
• run_script_isic.py for ISIC

In each script, we have provided configurations to run Marvell, max_norm, iso (isotropic gaussian) and no_noise (no perturbation) The corresponding command line option for each of the perturbation methods are:

• Marvell: --noise_layer_function sumKL
• max_norm: --noise_layer_function expectation
• iso: --noise_layer_function white_gaussian
• no_noise: use --noise_layer_function white_gaussian --ratio 0.0

Visualization

• We have provided diff_methods_tradeoff_viz_save_memory-*.ipynb to visualize the tradeoff results in tensorboard logs.