In april 2020 the approach by Levine et al. to track downloaders was proven to be wrong: Their false positives rate is wrong, their math is wrong, and their model is wrong. Therefore results you get when using their method are false. Yet this approach is still used two years later to get warrants against Freenet users.
The Levine group themselves changed their method in response (they posted this to conference proceedings, not peer-reviewed, and unlikely to pass peer-review because it contains claims that are obviously false), but they still use the 2017 method for arguments in court.
Therefore we are now providing the arguments that disprove it directly on the Freenet Project website to ensure that they are found when innocents are caught in that dragnet.
Wrong false positives rate
The core pillar of the detection they name is their claim of a 2.3% false positives rate. But this claim is wrong, because they only reach it through many false assumptions:
They ignore that friend-of-a-friend routing breaks their metric when
- an intermediary node, or
- the observing node has many connections.
which is not the rare case but the normal case.
They assume that they only get a false positive, if a request for a given file reached them with both HTL 18 and HTL 16 or with both HTL 17 and HTL 16. But the routing algorithm within Freenet causes them to almost always receive requests from a given node over the same route. So they will have the same HTL, regardless of the number of hops from the source to an observer. Therefore Their 2.3% false positives rate contains a mixture of
the probability of two people requesting the file in the same interval and
the rate of routing-changes within Freenet (for example because a node on the path went offline). If a request from a given peer is received both from HTL 17 and from HTL 16 (or 18 and 16) then routing changed, otherwise this should not happen.
Their false positives rate when measuring with only one node is therefore meaningless.
In addition their math is wrong:
We construct a model by assuming that each request the downloader makes is sent to exactly one of its peers, and that the selection of that peer is made uniformly at random.
This does not take friend of a friend routing into account. Therefore their math is wrong: It does not match the actual selection of peers, so the results are meaningless for the actual Freenet.
They expect even share peer selection, but Freenet does not use even share.
And their model of routing in Freenet, required to understand how their measurement works and what it observes, is wrong:
a simple expected fraction of 1/degree for the adjacent and (1/degree)² for the two-hop case.
This does not take the degree of the measuring node into account, therefore it is not a model of routing in Freenet.
This is in addition to the wrong math, an error they repeat here: Freenet does not use even share. Assuming scaling by inverse degree is wrong.
Their false positives rate is wrong, their math is wrong, and their model is wrong. Therefore results you get when using their method are false.
To somewhat save the grace for the Levine-team: They at least tried to actually measure the false positives rate. They did it wrong and drew false conclusions, and that they tried doesn’t make it right and it doesn’t excuse persecuting people based on their flawed reasoning, but at least they tried.
That the Levine-team did not contact Freenet developers prior to publication is inexcusable, though. It’s like publishing a paper based on evaluations of particle beams from CERN without ever talking to someone from CERN. The minimum of due dilligence would be to write an email to press -at- freenetproject.org saying “Hi, we found a method to track Freenet downloaders and drafted a paper based on that. Could you have a look to see whether we missed something?”
If you want to know the actual requirements for calculating a false positives rate in Freenet, read the article Statistical results without false positives check are most likely wrong.
While the Levine group has a false positives check, their check is wrong. They measured the wrong metric. We have explained patiently where the Levine group made mistakes. It is hard to understand when assuming scientific integrity that they still claim in court that their 2017 method is robust even after they changed their approach themselves.