Mathias Oliver Valdbjørn Jørgensen will give a talk about his MSc thesis work on automatic attacker generation for probabilistic privacy risk analysis. Details below.
Mathias Oliver Valdbjørn Jørgensen, MSc student, ITU.
Privugger-AG: Automatic Attacker Generation for Probabilistic Privacy Risk Analysis
I consider the problem of generating attackers in order to measure the leakage of a program disclosing data about individuals. I focus on these programs as black-box function and on leakage measurements of ratio scale type. My goal is to create a tool which consistently can automatically generate probabilistic attackers and ensure that the found attackers maximises a leakage based on a specific black-box function.
The current state of the art for analysing leakage is a method called Privug, which estimates leakage by reinterpreting a regular program probabilistically. It uses probabilistic distribution to model both individuals to be disclosed and Bayesian Inference to estimate the output. This allows information theoretical analysis of the program leakage. It suffers however, significantly by its choice of probabilistic distribution, which currently is done manually.
In this paper, I take advantage of the research within Privug to make the foundation of how to measure leakage probabilistically. I show that by automatically scoping the domain of possible distribution, based on the signature of a disclosure program, we can maximise the leakage based on a measurement. The maximization is done using Bayesian Optimisation, which have shown to converge to the global maximum with few executions of the black-box function. I evaluate this approach to other Derivative Free Optimization techniques. I demonstrate the reliability of the tool on both synthetic and real world libraries for disclosure of data. The tool can be used on any leakage measures and I show it to converge on leakage measurement such as Entropy, Mutual Information and Bayes Risk.