Task-aware Privacy Preservation for Multi-dimensional Data

This is the code release for our ICML 2022 paper "Task-aware Privacy Preservation for Multi-dimensional Data".

arXiv: https://arxiv.org/abs/2110.02329

Setup

• export LDP_CODESIGN_DIR in your bashrc to be the path of this repo

  • example export LDP_CODESIGN_DIR=~/Desktop/task_aware_privacy/

• have python3 installed. My version is 3.7.4

• have other python packages installed, including torch, numpy, pandas, etc

Models

There are 5 subdirectories for evaluation under ./codesign/:

• eval_theoretical computes some theoretical results under Assumption 1;

• eval_household, eval_valuation and eval_cancer correspond to the hourly household power consumption, real estate valuation, breast cancer detection applications, respectively; eval_mnist corresponds to an evaluation on high-dimensional image dataset -- MNIST.

The parameters should be specified in the get_model_specific_info.sh first.

Approaches

There are 3 approaches: codesign, sep_design, and benchmark, corresponding to task-aware, privacy-agnostic and task-agnostic approaches mentioned in our paper.

Theoretical Result

cd ./codesign/eval_theoretical

# Compute the theoretical losses
./laplace_theoretical_loss.sh

# Plot loss under different privacy budgets
./laplace_loss_privacy_plot.sh 

Simulation

• [linear model and l2-loss] For household application:

cd ./codesign/eval_household

# Read raw data
./household_data_generation.sh

# Co-design, sep-design, benchmark
./household_codesign_computation.sh 
./household_sep_design_computation.sh
./household_benchmark_computation.sh

# Plot loss under different privacy budgets / MSE for each x_i
./household_loss_privacy_plot.sh
./household_mse_plot.sh 

Caveat: due to randomness running the algorithms twice may produce slightly different results.

• [General Settings, regression] For valuation application:

cd ./codesign/eval_valuation

# Read raw data and train regressor
./valuation_data_generation.sh 
./valuation_train_regressor.sh

# Co-design, sep-design, benchmark
./valuation_codesign_train_encoder.sh
./valuation_sep_design_train_encoder.sh
./valuation_benchmark_train_encoder.sh

# Plot loss under different privacy budgets
./valuation_loss_privacy_plot.sh

• [General Settings, classification] For cancer application:

cd ./codesign/eval_cancer

# Read raw data and train classifier
./cancer_data_generation.sh
./cancer_train_classifier.sh 

# Co-design, sep-design, benchmark
./cancer_codesign_train_encoder.sh
./cancer_sep_design_train_encoder.sh 
./cancer_benchmark_train_encoder.sh

# Plot loss under different privacy budgets
./cancer_loss_privacy_plot.sh

• [General Settings, classification, minibatch] For mnist application:

cd ./codesign/eval_mnist

# Read raw data and train classifier
./mnist_data_generation.sh
./mnist_train_classifier.sh

# Co-design, sep-design, benchmark
./mnist_codesign_train_encoder.sh
./mnist_sep_design_train_encoder.sh
./mnist_benchmark_train_encoder.sh

# Plot loss under different privacy budgets
./mnist_loss_privacy_plot.sh

We also have a few scripts that visualize data/result in other ways, including (take cancer as an example):

# The distribution of data for each x_i and each label y (only for classification) 
./cancer_data_plot.sh

# Train loss evolution (for all the applications)
./cancer_train_losses_plot.sh

Tuning Hyperparameters and Models

The number of back-propagation steps in each epoch can be specified in ./codesign/ML_functions/LDP_encoder_train.py (or LDP_encoder_train_minibatch.py).

The encoder/decoder, classifier and regressor model can be further specified in the corresponding files under codesign/models. (For our paper, to repeat our experiment given in the appendix, the activation functions of the encoder/decoder model shall be changed manually.)