Skip to content

Latest commit

 

History

History

testing

DP Stochastic Tester

This is a framework that attempts to falsify the DP predicate for a given algorithm F over a set of datasets. For a given dataset D, we check the predicate by considering all adjacent datasets in the powerset of D, generating the probability distributions of the over the output of algorithm F. To give us more confidence that an algorithm is DP, the typical usage is to choose a diverse set of datasets. Note that since this problem is semidecidable, this method does not prove that an algorithm is DP, but can determine if an algorithm is not DP. Details can be found in section 5.3 of our paper.

How to Use

stochastic_tester_test.cc contains typical usage examples. To run the tests, use

cd cc
bazel test testing:stochastic_tester_test

We also run through a simple example here. First, create a Halton sequence.

auto sequence = absl::make_unique<HaltonSequence<double>>(
      /*dimension=*/3, /*sorted_only=*/true, /*scale=*/1, /*offset=*/.5);

Here sequence is an object that can be used to generate a determinisitic sequence of uniform random input sets for our algorithms that are spread out "evenly". The dimension is the size of the input. The scale and offset above imply that the input points are in the range [-0.5, 0.5]. Next, create the algorithm you want to test.

std::unique_ptr<Count<double>> algorithm =
      Count<double>::Builder()
          .SetLaplaceMechanism(
              absl::make_unique<test_utils::SeededLaplaceMechanism::Builder>())
          .SetEpsilon(.1)
          .Build()
          .ValueOrDie();

To avoid flakiness in our test, we use the deterministic SeededLaplaceMechanism. Finally, create an instance of the tester.

StochasticTester<double, int64> tester(
    std::move(algorithm), std::move(sequence),
    /*num_datasets=*/500, /*num_samples_per_histogram=*/20000);

The num_datasets specifies the number of inputs sets we want to check for whether they violate the DP predicate. For each input set, the num_samples_per_histogram specifies how many runs of the algorithm the tester will do to generate output distribution histograms. The higher these numbers, the higher the likelihood to catch any DP violations should they exist. However, increasing the numbers also makes the test slower. Finally, check the status of running the tester.

bool test_passes = tester.Run();

If the result is true, then the tester didn't detect a DP violation. Otherwise, the tester will log additional output to help your debugging.