Skip to Main content Skip to Navigation

Algorithms and techniques for bot detection in social networks

Maksim Kalameyets 1 
Abstract : In this thesis, we propose machine learning techniques to detecting and characterizing malicious bots in social networks. The novelty of these techniques is that only interaction patterns of friends' of analysed accounts are used as the source data to detect bots. The proposed techniques have a number of novel advantages. There is no need to download a large amount of text and media data, which are highly language-dependent. Furthermore, it allows one to detect bots that are hidden by privacy settings or blocked, to detect cam- ouflages bots that mimic real people, to detect a group of bots, and to estimate quality and price of a bot. In the developed solution, we propose to extract the input data for the analysis in form of a social graphs, using a hierarchical social network model. After, to construct features from this graph we use statistical methods, graph algorithms, and methods that analyze graph embedding. And finally, we make a decision using a random forest model or a neural network. Based on this schema we propose 4 techniques, that allows one to perform the full cycle attack detection pipeline - 2 techniques for bot detection: individual bot detection, and group bot detection; and 2 techniques for characterization of bots: estimation of bot quality, and estimation of bot price. The thesis also presents experiments that evaluate the proposed solutions on the example of bot detection in VKontakte social network. For this, we developed the software prototype that implements the entire chain of analysis - from data collection to decision making. And in order to train the models, we collected the data about bots with different quality, price and camouflage strategies directly from the bot sellers. The study showed that using only information about the graphs of friends it is possible to recognize and characterize bots with high efficiency (AUC - ROC ˜ 0.9). At the same time, the proposed solution is quite resistant to the emergence of new types of bots, and to bots of various types - from automatically generated and hacked accounts to users that perform malicious activity for money.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Monday, February 7, 2022 - 6:49:13 PM
Last modification on : Monday, July 4, 2022 - 9:17:17 AM
Long-term archiving on: : Sunday, May 8, 2022 - 7:26:37 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03560882, version 1


Maksim Kalameyets. Algorithms and techniques for bot detection in social networks. Library and information sciences. Université Paul Sabatier - Toulouse III; ITMO University, 2021. English. ⟨NNT : 2021TOU30097⟩. ⟨tel-03560882⟩



Record views


Files downloads