Speaker: Dr Gábor Gulyás, Inria
UCL Contact: Jonathan Bootle (Visitors from outside UCL please email in advance).
Date/Time: 31 Mar 16, 16:00 - 17:00
Venue: Roberts 110
The first demonstration in 2009 showed that structure is enough to map large proportions of two different social networks efficiently with a relatively small error rate. This poses several risks to users' privacy, as formerly unrelated datasets could be linked together (e.g., phone call and email communication networks), and anonymity could be broken in released datasets by using public data with known identities.
This talk addresses two issues: discussing advanced attacks to discover attacker limits and how we can further automatize these attacks with machine learning. Therefore, in the first part, we will discuss two attacks that have several salient properties compared to existing works, such as almost error-free results (with false positive rates well below 1%) and particularly high rates of correctly re-identified nodes. Then we will discuss work-in-progress results on how machine learning can be applied to achieve large-scale re-identification efficiently.
Dr Gábor Gulyás
Gábor György Gulyás is a postdoctoral researcher in the Privatics team, at Inria. He currently works at the intersection of data privacy and machine learning, but he is also interested in some specific areas, such as web privacy, re-identification attacks and privacy in social networks. He received his Ph.D. in 2015 from the Budapest University of Technology and Economics (BME), Hungary, where he was working on re-identification in social networks in the CrySyS Lab. He received the degree of MSc in a specialization on computer security at BME, Hungary.