Characterizing the Use of Images in State-Sponsored Information Warfare Operations by Russian Trolls on Twitter

Savvas Zannettou, Tristan Caulfield, Barry Bradlyn, Emiliano De Cristofaro, Gianluca Stringhini, Jeremy BlackburnICWSM 2020 · OpenReview

TL;DR

This paper analyzes 1.8 million images shared by Russian state-sponsored troll accounts on Twitter to understand their content, targets, and influence. The study reveals that these trolls strategically use politically charged images to sow discord and push agendas, demonstrating a significant impact on online discourse, particularly concerning political content.

What Problem Does This Paper Address?

In today's digital age, social media platforms have become battlegrounds for information warfare, where state-sponsored actors use various tactics to manipulate public opinion and push specific agendas. While previous research has focused on textual content and the general behavior of these "troll" accounts, the specific role and impact of images in these operations have been less understood. Images, with their inherent visual appeal and ability to convey complex messages quickly, are powerful tools for engagement and persuasion. This paper aims to fill this gap by providing a comprehensive analysis of the images used by Russian state-sponsored trolls on Twitter, shedding light on their content, the entities they target, and their effectiveness in disseminating information across various online communities.

Methods

The researchers developed an image analysis pipeline building upon existing work. This pipeline involves:

  • Perceptual Hashing (pHash): Used to group visually similar images, allowing for efficient clustering and identification of duplicate or near-duplicate content.
  • Clustering: Grouping images based on their pHashes to identify sets of visually related content.
  • Google Cloud Vision API: This state-of-the-art computer vision tool was employed to annotate images. It provides:
  • Entity Recognition: Identifying objects, people, and concepts within an image (e.g., "Donald Trump," "Russia").
  • Web Detection: Finding URLs where the same or similar images have appeared online, providing context about their dissemination.
  • Hawkes Processes: A statistical modeling technique used to quantify the influence and interplay between different online communities (Twitter, Reddit, Gab, 4chan's /pol/) in the spread of images. This method helps estimate the probability that an image's appearance in one community was triggered by its presence in another.

Datasets

The primary dataset used in this study consists of:

  • 1.8 million images extracted from over 9 million tweets posted by 3.6K Russian state-sponsored accounts identified by Twitter. This dataset was released by Twitter in October 2018 as part of their efforts to combat state-sponsored propaganda.
  • A Cross-Platform Dataset containing 160 million pHashes and image URLs from Twitter, Reddit, 4chan's /pol/, and Gab, spanning July 2016 to July 2017. This dataset was used to track image occurrences across different platforms.

Key Insights and Ideas

  • Images as a Strategic Tool: Russian trolls don't just use text; they strategically employ images, often politically charged or meme-based, to increase engagement and credibility.
  • Coincidence with Real-World Events: Image posting activity by trolls often peaked around significant real-world events, such as the Charlottesville rally, suggesting a deliberate effort to exploit and amplify societal divisions.
  • Targeting and Themes: The content of the images primarily focused on Russia, Ukraine, and the USA, indicating specific geopolitical interests.
  • Influence Beyond Twitter: The study highlights how images shared by trolls appear across various mainstream and alternative Web communities, demonstrating a broad reach.
  • Political Content Dominance: Trolls were more effective in disseminating politics-related imagery compared to other types of images.
  • Visual Cues Drive Sharing: The research reinforces the idea that users often rely on visual cues (like thumbnails) to decide what to share, making image-based disinformation particularly potent.

Key Results

  • Extensive Image Use: 19% of tweets from Russian trolls contained at least one image, with a significant portion being explicitly posted rather than automatically generated thumbnails.
  • Temporal Shift: Image usage increased significantly after 2016, coinciding with a shift in targeting towards specific foreign countries like the US.
  • Peak Activity: A notable peak in image dissemination occurred around the Charlottesville rally in August 2017.
  • Influence on Online Communities: Russian state-sponsored accounts were found to be particularly influential in spreading political imagery to communities like Gab and The_Donald subreddit.
  • Efficiency Comparison: While trolls were more influential in spreading political images than general images, they were slightly less influential in spreading news via URLs compared to images.
  • Political Image Effectiveness: Russian trolls were more efficient in spreading political images (both Republican and Democratic) compared to their general image dissemination.

Limitations and Future Work

The authors acknowledge several limitations:

  • Closed-System API: Reliance on Google's Cloud Vision API, a closed-system with an unknown methodology for entity extraction, introduces potential biases. However, a small-scale manual evaluation suggested acceptable performance.
  • Image-Only Analysis: The study primarily focuses on images, potentially missing crucial context from accompanying text, sentiment, or toxicity analysis.
  • Dataset Provenance: The dataset of identified troll accounts was provided by Twitter, and the methodology for this identification is not publicly available, raising questions about potential false positives.

Future work suggested includes:

  • Detecting Doctored Images: Developing methods to identify possibly doctored images.
  • Domain Control: Building on the work to detect domains controlled by state-sponsored actors for pushing specific narratives.
  • Broader Application: Applying the developed tools and techniques to study image sharing by other state-sponsored actors, bots, or coordinated campaigns, including newly released datasets from Venezuela and Bangladesh.
  • Interdisciplinary Impact: The findings are intended to inform policymakers, law enforcement, military personnel, and researchers studying information warfare and its societal impact.

Summary updated

BibTeX
@inproceedings{zannettou2020characterizing,
  author = {Savvas Zannettou and Tristan Caulfield and Barry Bradlyn and Emiliano De Cristofaro and Gianluca Stringhini and Jeremy Blackburn},
  title = {{Characterizing the Use of Images in State-Sponsored Information Warfare Operations by Russian Trolls on Twitter}},
  booktitle = {ICWSM},
  year = {2020},
  url = {https://openreview.net/forum?id=tyUKF0FUMuB},
}
Figure