Activists just got another reason to worry about what spooks might be able to learn about them, with boffins demonstrating that a decent traffic fingerprint can tell an attacker what's going on, even if an app is defended by encryption.
The researchers from the Universities of Padua and Rome have found that for activities like posting messages on a friend's Facebook wall, browsing a profile on a social network, or sending an e-mail, there's no need to decrypt an encrypted data flow.
In this paper at Arxiv, the researchers note that even in the hands of a knowledgable user there are opportunities for “malicious adversaries willing to trace people … the adversary can still infer a significant amount of information from the properly encrypted traffic.”
For example, if an attacker “knows” what a Twitter post's traffic typically looks like even if encrypted, and can observe both the traffic from the phone and the timing of a post appearing on Twitter, it is possible to de-anonymise the posts and associate them with the phone (and therefore the user), they assert.
The attack scenarios tested in the paper include identifying an activist whose traffic is encrypted, by correlating the inferred activity coming from a device with the time that posts appear on social networks; identifying communications between users on a similar basis; or building a behavioural profile of a target.
Inputs to the analysis included IP addresses and ports (which aren't encrypted, even if user “content” is), along with size, direction (incoming or outgoing) and timing of the communications, along with machine learning techniques using samples of traffic from apps like Facebook, Gmail and Twitter to act as seeds for their algorithms.
Those algorithms plus the visible TCP/IP data provide a data set that allows user actions – read or send e-mail, post on Twitter, post on Facebook and so on – to be classified without looking inside the packets, they write: “Our classification method is based on the detection of the sets of flows that are distinctive of a particular user action.”
They also used different accounts to train their algorithms from the accounts used in their test set, to “assure that the results of the classification do not depend on the specific accounts that have been analysed”. The study claims the tests show a classification accuracy of different actions between 91 per cent and 100 per cent.
“A powerful attacker such as a Government, could use these insights in order to de-anonimize user actions that may be of particular interest” is the study's glum conclusion.
Note: While such work has been performed on fixed networks, the boffins note that the techniques in prior work could not easily be moved to the mobile sphere. ®