Introduction

PrivaSeq is a toolbase for quantification and analysis of the individual characterizing information leakage, which can be used to link phenotype datasets to genotype datasets and reveal sensitive information in linking attacks.

For technical details please refer to: "Quantification of private information leakage from phenotype-genotype data: linking attacks", Nature Methods, 2016.

The motivation for analysis of linking attacks is motivated by the recent surge of high dimensional phenotyping datasets, which are served with personal information after being "anonymized".

It is becoming clear that we need to proactively evaluate the risks associated with how well an adversary can link the phenotype datasets to the genotype datasets and to other sensitive information.

PrivaSeq provides several tools that can be utilized for estimating the information in the phenotype datasets that can be used to link them to genotype datasets. These links are mediated through quantitative trait loci (QTL) datasets, which enable prediction of genotypes from phenotypic information.

The quantifications can be used to direct the data publishing mechanisms to identify the primary sources genotypic information "leakage" in the phenotype datasets and control how much information leaks. These can be used for risk assessment.

Analysis Software

We are in the process of building the github repository. In the meantime, you can download the leakage quantification and vulnerable fraction computation code from here. The directory contains a README.txt with file formats that the code uses and an example for running the analysis in the manuscript. Currently, only the leakage from eQTL datasets are considered. The code can evaluate 3 sources of leakage:

• Per eQTL Information Leakage: For each eQTL, code computes the amount of information that is leaked for the given expression and genotype dataset. This leakage is an overall estimate of how much information this eQTL would leak in general. High information leaking eQTLs can be excluded from queried databases or from the published data.
• Cumulative leakage of Information over sorted eQTL lists: After per eQTL analysis, one should also evaluate, how much predictability and leakage does a list of eQTLs convey to an attacker. This can be used to estimate risks associated with releasing the list of eQTLs
• Extremity based linking attack: The linking attack can be used when genotype and phenotype datasets are to be published/served. The vulnerable individuals can be excluded from the datasets.

The outputs from each analysis are explained in more detail in the README.txt file.

We are working on extending the code to easily handle other data formats and other types of QTLs so that the leakage estimates can be computed for other data types.

Arif Harmanci, arif.harmanci@yale.edu.