Feature Selection Techniques for Enhancing Sybil Accounts Detection in Online Social Networks

Abstract

The 24X7 affordable availability of Online Social Network (OSNs) is attracting more and more users to get connected to these social media. Fake users and spambots too are vying for claiming their illegitimate share on these OSNs. The OSNs, in the process, are also generating large datasets pertaining to details of all these types of users which normally contain huge sets of features both, relevant and irrelevant. These datasets can be used of for initiating some preventive measures against various attacks such as Sybil attack. Feature selection process applied appropriately on the dataset(s) helps in improving the efficiency and overall performance of predictive models. Based on the necessary selection criterion, three main categories of feature selection techniques developed over the time are: filter, wrapper, and embedded methods. There is a need to study and compare the performances of these feature selection techniques on the same sets of datasets under various prediction models. Thus, in this paper, I have conducted such experiments on twitter datasets containing accounts of real users, fake followers and spambots. The filter methods used for our experimentation are univariate and correlation matrix with heatmap. The Genetic Algorithm (GA) and Recursive Feature Elimination (RFE) techniques as parts of wrapper methods have been used separately. For embedded method, we have implemented Lasso regression. The results have shown that the filter methods provide the best results for datasets involving fake followers and the wrapper methods provide better performance in case of spambot datasets.