Detecting Malicious Nodes in Social Networks

Recently, there has been increasing interest to identify nodes in social networks that behave in a malicious way. Especially, Social Bots pose a major problem in many networks.

Methods for identifying these nodes can focus on three aspects:

Content-based methods concentrate on the content generated by the node and use e.g. NLP methods for classification.

Structural approaches analyze the node’s neighbourhood  in the social network for finding differences to ‘normal’ nodes.

Behavioral methods regard the dynamic aspects of nodes in order to detect abnormalities.

For each of these methods, specific machine learning methods have been used, like classification based on feature vectors for content, graph-oriented approaches for structure and Markov models for the dynamic aspects. However, this diversity of methods leads to the problem that the consideration of more than one aspect is not easily possible.

In this thesis, approaches for integration of these methods will be investigated:

Loose: Integration combining 3 specialized classifiers with a meta-learner.

Pipelining: The output of a method forms the input for other method(s).

Close: Integration using a single classifier suitable for considering all aspects.

For more information, please contact Prof. Norbert Fuhr