|
@DocEdge85 | |||||
|
(Thread) Here’s my new preprint, with @Graham_Coop, on genetic privacy in genealogy databases that allow user uploads. (1/n)
biorxiv.org/content/10.110…
|
||||||
|
||||||
|
Doc Edge
@DocEdge85
|
22. lis |
|
We also wrote an explainer that’s posted on Graham’s blog (2/n)
gcbias.org/2019/10/22/pap…
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
In this paper, we’re thinking about direct-to-consumer (DTC) genetic genealogy services that allow users to upload their own genetic datasets. Examples include GEDmatch, @FamilyTreeDNA, @MyHeritage, and @LivingDNA. (3/n) pic.twitter.com/r088F2Wy3k
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
There are advantages to these upload features: users find more biological relatives without having to be re-genotyped (or pay for new genotyping), and the services grow their databases, which lets them connect more people. (4/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
But there are always risks to sharing genetic information. For example, if a relative matches with you in a particular chromosomal location, then your relative learns something about your genotypes in that location, as pointed out by Leah Larkin (5/n)
thednageek.com/cystic-fibrosi…
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Our point in this paper is that if a user can upload many datasets, there are ways to learn genotypes that are supposed to be private. Depending on how the DTC service is configured, the majority of genotypes in the database can be revealed with a few hundred uploads. (6/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
We explore three ways of doing this. We call the first “tiling,” and to do it, an adversary would just upload many genomes (e.g. publicly available ones) and keep track of their chromosomal matches with genomes in the database. (7/n) pic.twitter.com/0a0nl4LbN2
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
We find that tiling works pretty well in a set of publicly available European genomes. (8/n) pic.twitter.com/BiHzErSmTc
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
The second approach is “probing,” and you could use it to find genotypes at a specific genomic location even if the DTC service doesn’t have a chromosome browser. (9/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
The idea is to fill in most of the uploaded genome with fake genotypes designed to match no one, so any people returned as matches will match the uploaded dataset at the site of interest. (10/n) pic.twitter.com/rxNB1Znaf6
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
The third approach is called “baiting,” and it takes advantage of methods for identifying relatives that rely on unphased genetic data. (That is, genetic data where we don’t know which alleles go together on the same copy of a chromosome.) (11/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
You can identify relatives in unphased data by looking for long regions where two people are never homozygous for opposite alleles. (12/n)
journals.plos.org/plosone/articl…
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Though no longer cutting-edge, this method scales to big datasets. One big problem is that it’s easy to trick---upload a fake dataset that’s heterozygous at every site, and it will appear to match everyone everywhere, as if it were the parent of everyone in the database. (13/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
You can use this to reveal genotypes at specified locations. Upload a dataset that’s mostly het, but put in homozygous genotypes at the sites you want to know. If a person in the database has the opposite homozygous genotype, the inferred related segment will break. (14/n) pic.twitter.com/n8QWh95TGP
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Under the simplest version of this kind of relative-finding algorithm, we estimate that you could reveal enough genotypes to impute genome-wide genotypes of everybody in the database with 97%+ accuracy after about a hundred uploads. (15/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
And if you just care about one SNP---such as one that reveals APOE4 alleles and thus a lot of information about Alzheimer's risk, for example---you could get the genotypes of everyone in the database w/ 2 uploads of fake data. (16/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Luckily, all the attacks we describe are preventable, or at least can be made inefficient, if DTC services use a subset of the countermeasures we describe (17/n) pic.twitter.com/fWho3Ux7Oj
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Some countermeasures are easy, like only returning information about long chromosomal matches. Others, like requiring cryptographic signatures suggested by @erlichya @itsikp @ShaiCarmi last year, are harder to implement but very effective (18/n)
science.sciencemag.org/content/362/64…
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
We wrote to all the DTC genealogy services we know of that allow uploads 90 days ago to share these methods and the countermeasures we recommend. They all wrote back to us, and some of them told us that some of these countermeasures are already in place. (19/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
We’d encourage all services that might potentially be affected by these kinds of attacks to share the countermeasures they have in place publicly. (20/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
Genetic genealogy can be an amazing, empowering thing for people who want to find their biological relatives, including folks who wouldn’t be able to find them otherwise. Our goal (shared w/ the companies offering these services) is that people be able to do this safely. (21/n)
|
||
|
|
||
|
Doc Edge
@DocEdge85
|
22. lis |
|
And of course, this is all part of a larger conversation about how we as a society want our genetic information used and want our genetic privacy protected. (Add it to a list of reasons to @extendGINA) (end)
|
||
|
|
||