The phenomenon of fake accounts on social networks a widespread problem, despite the techniques implemented by the platforms to prevent attackers from registering profiles with which to perpetrate abuse (spam, etc.). After passing the initial check, a fake account has to do nothing more create a social network by sending friend requests, until you have a sufficient number of contacts to expose to the various offenses.
Facebook obviously on the front lines to combat this phenomenon and in the past few hours he has published research in which he explains how can identify fake accounts who manage to overcome existing controls and who still do not have enough connections with real users to commit abuse. Everything revolves around a algorithm called SybilEdge which processes how users add friends to their network. By looking at the friend requests made by the fake accounts and how many of them have been accepted or rejected, SybilEdge now can accurately detect fake accounts with only 20 friend requests.
How is it possible? Facebook has been watching the actions of real and fake accounts for some time, finding that the two groups of users "differed significantly in both the selection of potential friends and the response of those targets to friend requests." The requests for fake accounts were rejected more often than those of real users. In addition, fake accounts were often more careful in choosing their goals, probably to maximize the possibility that their request would be accepted.
Facebook he thus created a behavioral champion with whom he trained SybilEdge, segmenting users into two groups: those most likely to accept friend requests from real accounts and those most willing to accept requests from fake accounts. If the former rejects a request, this behavior signals that the requester is a legitimate user. On the other hand, if users who accept multiple fake requests accept one, it means that the applicant is probably a fake.
SybilEdge operates in two phases. In the first, the algorithm is trained by observing the samples mentioned over time, while in the second, it uses classifiers that refer to reports based on real abuse. This training phase provides the model with all the parameters necessary to carry out in real time – for each friend request and response – an update on the probability that the applicant is a fake.
Facebook claims that SybilEdge has proven to detect fake accounts with an accuracy of over 90% with 15 or less average friend requests and an 80% accuracy in detecting fakes with 5 friend requests. Performance also does not degrade with increasing demands – over 45.
"SybilEdge helps us identify those who perpetrate abuse quickly and in a way that can be explained and analyzed. In the near future, we plan to explore other ways to further speed up detection of fake accounts and make better decisions even faster than SybilEdge. We plan to achieve this by mixing models based on characteristics and behavior, "concluded Facebook.