The logistic regression example provides a fast and scalable implementation of SEAL CKKS HE scheme based logistic regression. It will be built whenever SEAL is enabled as part of the HE Toolkit build.
This example is also capable of running linear regression instead of logistic
regression. Logistic regression can be achieved by
"wrapping" a
multiple linear
regression
model with a sigmoid
function. The sigmoid function
can be skipped by using the --linear_regression option, effectively running a
linear regression.
python3 >= 3.5
virtualenv
Packages needed for generating the synthetic datasets are automatically
installed in a virtual environment within
$HOME/.hekit/components/examples/logistic-regression/build/datasets/.
Assuming the user used hekit to build the example program using the provided
recipe file recipes/examples.toml, it can be run by executing
cd $HOME/.hekit/components/examples/logistic-regression/build
./lr_test--data: Dataset name. Default is lrtest_mid.
There are four different synthetic datasets available for testing, which are automatically generated during build time.
| Name | Features | # Samples |
|---|---|---|
| lrtest_small | 40 | 500 |
| lrtest_mid | 80 | 2000 |
| lrtest_large | 120 | 10000 |
| lrtest_xlarge | 200 | 50000 |
--poly_modulus_degree: Polynomial modulus degree. Determines the
encoding slot count (half of the parameter) and encryption security level.
Default is 8192, and recommended size is {4096, 8192, 16384}, and
must be a power of 2, with full range of [1024, 32768].
--compare: Compare the HE logistic regression inference result with non-HE
inference for validation purposes. Default is false.
--data_plain: Run with the data as plaintext.
--model_plain: Run with the model as plaintext.
--linear_regression: Calculate linear regression instead of logistic regression.
--security_level: Security level. One of [0, 128, 192, 256].
--coeff_modulus: Coefficient modulus (list of primes). The bit-lengths of the primes to be generated.
--batch_size: Batch size. 0 = automatic (poly_modulus_degree / 2). Max = poly_modulus_degree / 2.
--scale: Scaling parameter defining precision.
There are two example data preparation ipython notebooks in datasets folder.