Author: Georgios Damaskinos ([email protected])
DP-SCD is a privacy-preserving optimization algorithm introduced in Differentially Private Stochastic Coordinate Descent.
- Python (3.5+)
pip install sklearn pandas seaborn wget sympy
The following reproduces a single (out of the total of 10 used) run for the result shown in Figure 1(a) (DP-SCD (ε=0.1)).
python driver.py --app RR --solver PSCD --max_iter 50 --lambda 1e-4 --eps 0.1 --lot_ratio 1000 --C 0.1 --dataset msd --dual --valid_size 0.25 --seed 1
-
Preprocessor: Loads the data and handles preprocessing. Information on how to obtain the data is embedded to the code.
-
Optimizer: Implements the various optimization algorithms, namely SCD, SGD, DP-SCD, DP-SGD, in an application-agnostic manner.
-
Applications (ridge regression, logistic regression, SVMs): Implement the application-specific code required by the Optimizer to perform training.
-
Driver: Tool for parallel (single machine) training and evaluation for a given set of hyperparameters.
-
Accountant: Tool for measuring the privacy loss based on the moments accountant.