Facebook engineers recently designed an algorithm, SybilEdge, which is used to identify fake accounts on the platform which might have escaped the anti-abuse filters at the time of registration but lack the number of friend requests to continue the abuse. In a blog post, while describing the algorithm, Facebook mentioned that the behavior and pattern of the account, the way it adds friends and expands its network, is analyzed after which its capability to attack other users is minimized.
The algorithm is capable of identifying fake Facebook accounts that have less than 20 friend requests and are hardly a week old or less. Platforms that are facing the spread of misinformation about Coronavirus can use SybilEdge in their favor.
According to an analysis by Reuters Institute for the Study of Journalism at the University of Oxford, around 33% of the people have come across misleading information about coronavirus on different social platforms like Facebook, Twitter, and YouTube.
The development team at Facebook while developing the SybilEdge figured out that to perform abusive activities, fake users have to add their targets in their friend list by finding and then sending friend requests to them. The Facebook internal study revealed that real accounts have different behavior and pattern when it comes to finding and adding friends and getting accepted. The fake requests are rejected more often than real users. Also, abusers try to increase the number of accepted requests by carefully choosing their friend request targets.
A corpus is created by Facebook to train the SybilEdge by dividing users into two sections; one who mostly accepts friend requests from real accounts and the other who often accept even the fake account requests. In case the users who accept requests from real users reject the request, it indicates that the requester is a real user. On the other hand, if the users who may accept requests from fake users accept the request, it hints that the requester could be a fake user.
There are two stages in which the algorithm works, first, it is trained by following the samples mentioned above and then it produces results depending upon the behavioral and content classifiers of Facebook that flag the users after an actual abuse.
In the training phase, the model with all the parameters is provided so the algorithm can work in real-time and update about the friend requests from potentially fake accounts.
According to Facebook, 90% of the detections by SybilEdge about fake accounts with less than 15 friend requests are correct. Also, when it comes to around 5 friend requests, almost 80% of the fake account detections by the algorithm are accurate. Even with the increased number of friend requests, above 45 or so, the algorithm still considerably perform better.
Facebook said SybilEdge allows it to identify fake accounts quickly in an easily analyzable way. The company plans to mix the feature-based and behavior-based models to detect fake accounts at a higher speed than SybilEdge and make many confident decisions about it.
Facebook is working to incorporate AI training technique, called self-supervised learning, which will improve the learning accuracy by using the unlabeled data along with some labeled data.
A 20% decrease in the abusive content on the platform was noticed within two years of the implication of deep entity classification (DEC) machine learning framework by Facebook. Facebook researchers carried out another experiment in which they trained the language understanding model that took only 80 hours to make much precise prediction than manually labeled data, which took 12,000 hours.