Method for detecting spammers and fake profiles in social networks

Patent metadata

Field	Value
Publication number	US-9659185-B2
Application number	US-201314386769-A
Country	US
Kind code	B2
Filing date	Mar 21, 2013
Priority date	Mar 22, 2012
Publication date	May 23, 2017
Grant date	May 23, 2017

Abstract

A method for protecting user privacy in an online social network, according to which negative examples of fake profiles and positive examples of legitimate profiles are chosen from the database of existing users of the social network. Then, a predetermined set of features is extracted for each chosen fake and legitimate profile, by dividing the friends or followers of the chosen examples to communities and analyzing the relationships of each node inside and between the communities. Classifiers that can detect other existing fake profiles according to their features are constructed and trained by using supervised learning.

First claim

The invention claimed is:

1. A method for protecting user privacy in an online social network, comprising the steps of:
a) providing a database storing profiles of existing users of said social network;
b) installing, on a non-transitory computer readable medium of a hardware processor, used to control a computer being connected to said social network, a community detection module, for splitting said social network into communities and extract features from each community;
c) choosing, by said community detection module, positive examples of legitimate profiles from said database;
d) choosing, by said community detection module, negative examples of fake profiles using simulation of fake profiles infiltration which is done automatically by said computer;
e) extracting, by said community detection module, a predetermined set of features for each chosen fake and legitimate profile by dividing the friends or followers of the chosen examples to communities and analyzing the relationships of each node inside and between said communities; and
f) constructing and training classifiers that can detect other existing fake profiles according to their features, using a supervised machine learning software module, installed on said computer.

2. A method according to claim 1 , wherein fake profiles in the social network is identified representing said social network as a directed graph.

3. A method according to claim 1 , wherein positive fake profile examples are obtained by using random friend requests.

4. A method according to claim 1 , wherein negative examples of fake profiles are obtained by randomly choosing legitimate profiles from said social network.

5. A method according to claim 1 , wherein the classifiers are trained for each of the positive and negative example generating a features vector for each user profile.

6. A method according to claim 5 , wherein features vectors are used as a training set for the fake profiles detection classifiers.

7. A method according to claim 1 , wherein a subset of the most likely fake profiles is manually evaluated, while using a set of randomly selected profiles as a control group.

8. A method according to claim 1 , wherein simulation of fake profiles infiltration in a directed social network is performed by:
a) representing the topology of said directed social network by a directed graph;
b) inserting new nodes to said graph, each of which representing a fake profile; and
c) inserting each fake profile into said graph by simulating sending a series of “follow” requests to random users on said directed social network, while limiting the number of friend requests that can be sent by each fake profile.

9. A method according to claim 5 , wherein a set of features is extracted for each user, said set consisting of:
a) the number of friends of said user;
b) the number of communities said user is connected to;
c) the number of connections between the friends of said user; and
d) the average number of friends inside each of the user's connected communities.

10. A method according to claim 1 , wherein fake profiles detection classifiers are constructed by:
a) automatically creating a subset of positive and negative examples with different sizes from each social network;
b) for each social network, removing users having a number of friends which is smaller than a predetermined value; and
c) randomly choosing negative examples from each social network.

Assignees

• B G Negev Tech And Applications Ltd

Inventors

• Elovici Yuval
• Fire Michael
• Katz Gilad

Classifications

• G06Q10/40

  Business processes related to social networking or social networking services · CPC title

• G06Q10/107 Primary

  Computer-aided management of electronic mailing [e-mailing] · CPC title

• G06F21/6245 Primary

  Protecting personal data, e.g. for financial or medical purposes · CPC title

• G06Q50/01

  Physics · mapped topic

• G06Q10/48

  using social graphs · CPC title