Abstract |
Online social media have become a primary source of information, achieving rapid content spreading across the globe. However, the content of online social media has raised
concerns about fake news spreading and user misinformation. Malicious users, known
as bot accounts, are able to exploit current online social media towards public opinion
manipulation via the spread of fake or biased information in a convincing manner.
Due to Twitter’s popularity and data access policies via API, it has become popular with
malicious actors. During the past decade researchers have shown cases of propaganda,
opinion manipulation, and influence during political events on Twitter. Measuring the
direct impact of bot activity on human decisions is a challenging task. However, research
has shown indirect correlations between bot activity and event outcomes, such as in the
2016 US Presidential Elections and stock market share prices.
In order to preserve online social media integrity, effective bot detection methods are
necessary. Current solutions lack effective bot detection in the general case and mostly
offer high performance on very specific case scenarios, making them impractical in realworld situations.
In this dissertation, we explore the evolution and challenges of the bot detection techniques, towards the creation of a novel general bot detection method (BotArtist). This
method minimizes the utilization of the Twitter API usage and outperforms existing stateof-the-art bot detection methods. Towards this goal, we successfully identified botnet
communities that share nearly identical content at nearly the same time. We meticulously
analyzed a variety of characteristics and evaluated the differences between normal and
bot accounts. Additionally, we applied the lessons we learned to develop machine learning
models capable of bot detection within specific discussion topics. We also conducted an
extensive evaluation of various account characteristics in terms of their predictive power
over short and long time periods.
This comprehensive approach led to the creation of BotArtist, currently the most versatile bot detection method. BotArtist has achieved remarkable results, with average F1
scores of 83.19 for specific dataset applications and 68.5 for general use cases. This represents an overall performance increase of nearly 10% compared to the current state-of-theart methods.
|