Source-linked AI summary
Disinformation Warfare: Understanding State-Sponsored Trolls on Twitter and Their Influence on the Web
Savvas Zannettou, Tristan Caulfield, Emiliano De Cristofaro, Michael Sirivianos, Gianluca Stringhini, Jeremy Blackburn
TL;DR
The paper addresses limited understanding of how state-sponsored trolls operate, what they disseminate, and how they influence the Web. It analyzes identified Russian troll tweets against random users and models cross-platform influence with Hawkes Processes. Trolls showed distinctive topics, identities, and platform behavior, but had minor influence on news diffusion except for RT content.
Problem
State-sponsored trolls’ operations, disseminated content, and influence across the Web have not been thoroughly studied.
Method
The study analyzes 27K tweets from 1K identified Russian trolls, compares them with random Twitter users, and applies Hawkes Processes to influence across Twitter, Reddit, and 4chan.
Results
Russian trolls discussed targeted political and event-related topics, changed identities, and behaved differently from random users, while their influence on news virality was minor except for RT links.
Takeaways & Limitations
The findings distinguish substantial troll activity and platform reach from generally limited ability to make news viral, with RT as a significant exception.
Takeaways & Limitations
The analyzed account list is incomplete because Twitter later identified over 1K additional troll accounts.
Abstract
from arXiv · showhide
Over the past couple of years, anecdotal evidence has emerged linking coordinated campaigns by state-sponsored actors with efforts to manipulate public opinion on the Web, often around major political events, through dedicated accounts, or "trolls." Although they are often involved in spreading disinformation on social media, there is little understanding of how these trolls operate, what type of content they disseminate, and most importantly their influence on the information ecosystem. In this paper, we shed light on these questions by analyzing 27K tweets posted by 1K Twitter users identified as having ties with Russia's Internet Research Agency and thus likely state-sponsored trolls. We compare their behavior to a random set of Twitter users, finding interesting differences in terms of the content they disseminate, the evolution of their account, as well as their general behavior and use of Twitter. Then, using Hawkes Processes, we quantify the influence that trolls had on the dissemination of news on social platforms like Twitter, Reddit, and 4chan. Overall, our findings indicate that Russian trolls managed to stay active for long periods of time and to reach a substantial number of Twitter users with their tweets. When looking at their ability of spreading news content and making it viral, however, we find that their effect on social platforms was minor, with the significant exception of news published by the Russian state-sponsored news outlet RT (Russia Today).
1 Introduction
The paper studies how Russian state-sponsored trolls operate, what they disseminate, and how much they influence the Web’s information ecosystem. It finds distinctive activity and identities, but generally minor influence on news diffusion except for RT links.
- Motivation: State-sponsored disinformation activity remains insufficiently understood, including trolls’ operations, content, and influence across the Web.Prior work focused mainly on bots, while human actors are also important in spreading false information on Twitter.
- Approach: 27K tweets from 1K Russian trolls were compared with random Twitter users to characterize their activity and quantify influence across Twitter, 4chan, and Reddit.The trolls were identified from 2.7K accounts released by the US Congress.
- Main findings: Russian trolls had very small influence on making news viral across social platforms, except for links from RT, which they effectively pushed.The exception concerns the Russian state-funded outlet Russia Today.
- Main findings: Trolls focused on specific world events, organizations such as ISIS, and political discussions involving Donald Trump and Hillary Clinton.The reported topics include the Charlottesville protests and political threads about the two candidates.
- Main findings: Trolls changed identities over time by deleting previous tweets and modifying screen names or profile information.The paper describes this behavior as resetting the profile.
- Main findings: Compared with random accounts, trolls reported locations concentrated in a few countries and mostly used Twitter’s Web Client rather than mobile versions.The countries highlighted are the USA, Germany, and Russia.
2 Background
The paper examines Twitter, Reddit, and 4chan as distinct social platforms that contribute to the Web’s information ecosystem. Their interaction models range from follower broadcasts and voting to anonymous imageboard discussions.
- Twitter: Twitter lets users broadcast short messages called tweets to followers, with hashtags for indexing and mentions for referring to other users.
- Reddit: Reddit is a news aggregator where users submit URLs and titles, while votes determine post popularity and ordering within topic-focused subreddits.
- 4chan: 4chan is an anonymous imageboard organized into topic-specific boards where users post images or text and reply to other posts.The study focuses on the Politically Incorrect board, /pol/.
3 Datasets
The dataset combines congressional identification of Russian troll accounts with sampled Twitter data and a matched random-user baseline. It contains 27K observed tweets from 1K trolls, while recognizing that the account list is incomplete.
- Russian trolls: The study starts from 2.7K Twitter accounts suspended for connections to Russia’s Internet Research Agency and released by the US Congress.The released list contains Twitter user IDs and screen names.
- Russian trolls: 27K tweets from 1K of those 2.7K accounts were retrieved between January 2016 and September 2017 from the 1% Twitter Streaming API dataset.
- Temporal analysis: Figure 1 reports temporal characteristics of tweets from Russian trolls and random Twitter users.
- Dataset limitation: The released account list is incomplete because Twitter later announced the discovery of over 1K additional troll accounts.The paper therefore treats the congressional list as valuable ground truth but not a complete census.
- Baseline dataset: The random-user baseline contains 1K users selected to match trolls’ distribution of average tweets per day, yielding 96K tweets over the same period.
4 Analysis
Russian trolls differed from baseline users in account management, self-presentation, posting tools, content, and activity patterns. Their tweets focused on specific political and world events, used distinctive media and hashtag practices, and reached substantial audiences.
- Account evolution: 71% of Russian troll accounts were created before 2016, with creation peaks shortly before the 2016 Republican National Convention and the 2017 Charlottesville rally.The timing may indicate coordinated activity around notable political events.
- Self-presentation: Trolls’ reported locations and timezones clustered in the USA, Russia, and selected European countries, unlike the baseline’s more globally diverse distribution.Two thirds of troll tweets appeared to use US timezones, while 18% used Russian timezones; the location pattern may reflect attempts to appear local.
- Content analysis: 66% of troll tweets contained no images, 32% contained exactly one image, and only 1.5% included a video; trolls also tended to post longer tweets.Baseline tweets had 73% with no images, 18% with exactly one image, and 6.4% with videos.
- Content analysis: Troll tweets covered specific events and political topics, used hashtags in 32% of tweets versus 10% for baseline users, and were more subjective and negative or neutral.Topics included Charlottesville, North Korea, Donald Trump, ISIS, and Islam; coordinated hashtag bursts included #Merkelmussbleiben and #IslamKills.
- Account evolution: 9% of troll accounts changed their screen name, sometimes up to four times, while trolls also deleted tweets in batches and substantially modified account profiles.These changes included adopting identities such as news outlets and altering follower and friend counts over time.
- Reach and activity: Russian trolls averaged 7K followers and 3K friends, while baseline users averaged 25K followers and 6K friends; troll tweets still reached at least 1K followers, peaking at 145K.These audiences provided a non-negligible potential reach for pushing specific narratives.
5 Influence Estimation
The paper estimates how URLs diffuse among trolls, Twitter users, Reddit, and 4chan using Hawkes processes. Troll influence is generally limited, but RT links are a notable exception.
- Influence model: Hawkes processes estimate connection strengths between Russian trolls, normal Twitter users, Reddit, and 4chan from URL-event sequences.Each weight represents the expected number of subsequent events in one group after an event in another.
- URL categories: RT URLs had higher background rates than other news URLs, indicating that they were more likely to occur spontaneously.The RT category contains only one news source.
- Connection structure: For /pol/, Reddit, and normal Twitter users, within-platform reposting had the greatest weights, whereas trolls connected more strongly to Twitter users than to themselves.The pattern suggests trolls’ Twitter accounts functioned as an avenue for disseminating information to normal Twitter users.
- RT dissemination: Trolls were more likely to retweet Russian state-sponsored URLs from normal Twitter users, while normal users were more likely to retweet those URLs from trolls.Most source-destination pairs were not statistically significant, but the trolls-to-Twitter-users comparison differed with p < 0.01.
- Estimated impact: For all URLs, trolls caused 0.01% of Twitter events, 0.93% on /pol/, and 0.62% on Reddit.These percentages measure destination-platform events attributed to source-platform events.
- Estimated impact: RT URLs produced approximately 2 times more impact on /pol/, 5 times more on Reddit, and 4 times more on Twitter than other news sources.The supplied passage reports these relative differences for troll influence on other platforms.
6 Related Work
Prior research examines opinion manipulation, political misinformation, and state-sponsored troll behavior across Web communities and platforms. These studies provide context for analyzing how Russian trolls operate and disseminate content.
- Scope: The paper situates its analysis within prior work on opinion manipulation, politically motivated disinformation, and state-sponsored troll behavior.These areas are explicitly identified as the scope of the related-work review.
- Opinion manipulation: Sockpuppets create multiple accounts to participate in Web communities and exhibit posting behavior different from benign users.Related work also reports that trolls can manipulate users’ opinions in forums.
- Political misinformation: Research on political misinformation has examined ideological interactions, false political information, and Twitter activity during elections.The cited work includes studies of right- and left-leaning communities during the 2010 US midterm elections.
- State-sponsored trolls: Studies of state-sponsored trolls use ground-truth datasets identified by social-network operators such as Facebook and Twitter.One cited study analyzes 10M posts by Russian and Iranian trolls to examine campaigns, targets, and changing behavior.
7 Conclusion
The paper finds that Russian trolls differed from random Twitter users in activity, content, identity changes, and follower growth, but had limited cross-platform influence except for RT news.
- 7 Conclusion: Russian trolls exhibited distinct Twitter behavior compared with random users, including active political dissemination, multiple identities, and efforts to increase followers.They analyzed activity over 21 months.
- 7 Conclusion: Hawkes Processes quantified troll influence across Twitter, Reddit, and /pol/.The model was applied to URL occurrences across these platforms.
- 7 Conclusion: Troll influence was not substantial on the other platforms, except for news published by the Russian state-sponsored outlet RT.The RT exception was significant.