DP-Opt: Identify High Differential Privacy Violation by Optimization

Ben Niu, Zejun Zhou, Yahong Chen, Jin Cao, Fenghua Li

Abstract: Differential privacy has become a golden standard for designing privacy-preserving randomized algorithms. However, such algorithms are subtle to design, as many of them are found to have incorrect privacy claim. To help identify this problem, one approach is designing disprovers to search for counterexamples that demonstrate high violation of claimed privacy level. In this paper, we present DP-Opt(mizer), a disprover that tries to search for counterexamples whose lower bounds on differential privacy exceed the claimed level of privacy guaranteed by the algorithm. We leverage the insights of counterexample construction proposed by the latest work, meanwhile resolve their limitations. We transform the search task into an improved optimization objective which takes into account the empirical error, then solve it with various off-the-shelf optimizers. An evaluation on a variety of both correct and incorrect algorithms illustrates that DP-Opt almost always produces stronger guarantees than the latest work up to a factor of 9.42, with runtime reduced by an average of $$19.2\%$$ .