The Many Shades of Anonymity: Characterizing Anonymous Social Media Content

Denzil Correa, Leandro Silva, Mainack Mondal, Fabrí­cio Benevenuto, Krishna P. GummadiAAAI 2021 · DOI

TL;DR

This paper introduces the concept of "anonymity sensitivity" to quantify how much users feel a piece of content should be anonymous, moving beyond a simple binary. They find that anonymous social media content, like that on Whisper, exhibits a wide spectrum of anonymity needs, unlike the generally low sensitivity of tweets. This nuanced understanding has implications for designing more privacy-aware online platforms.

What Problem Does This Paper Address?

The rise of anonymous social media platforms like Whisper and Secret presents a new frontier for online communication, where user identities are deliberately obscured. Unlike traditional platforms like Facebook and Twitter, where posts are tied to persistent user profiles, anonymous platforms lack this anchor. This raises questions about the nature of content shared in such environments: what kind of information do people feel comfortable sharing anonymously, and how does this differ from content shared on public, identity-linked platforms? Existing research often treats anonymity as a simple "on/off" switch, failing to capture the subtle differences in how sensitive users perceive different types of content to be. This paper aims to bridge this gap by characterizing anonymous content and exploring the nuances of anonymity needs.

Methods

The core innovation of this paper is the introduction and measurement of anonymity sensitivity (AS). This is defined as the probability that users would consider a piece of content to be anonymous. To measure this, the researchers conducted a large-scale human annotation study using Amazon Mechanical Turk (AMT). They presented 500 anonymous posts (whispers) and 500 public tweets to 100 AMT workers, asking them to label each as anonymous or non-anonymous. The AS score for a piece of content is then calculated as the fraction of annotators who labeled it as anonymous.

Beyond AS, the researchers also employed linguistic analysis using the Linguistic Inquiry and Word Count (LIWC) lexicon to identify distinct textual features of anonymous content. They then used these linguistic features to train a machine learning classifier to automatically distinguish between anonymous and non-anonymous posts.

Datasets

The study utilized two primary datasets:

  • Whisper Data: A large corpus of approximately 20.7 million "whispers" collected from the "Latest" section of the Whisper website between July 1 and November 17, 2014. This dataset included text, location, timestamp, number of favorites, and replies.
  • Twitter Data: A sample of 100,000 tweets collected from the Twitter Streaming API during the same period as the Whisper data.

For the human annotation experiment, a subset of 500 whispers and 500 tweets was used. For the linguistic analysis and automated detection, a larger sample of 100,000 whispers and 100,000 tweets was employed.

Key Insights and Ideas

  • Anonymity is a Spectrum, Not a Switch: The most significant finding is that anonymity sensitivity is not binary. Whispers exhibit a wide range of AS scores, from 0 to 1, indicating that different types of content have varying degrees of perceived anonymity needs. Tweets, in contrast, are overwhelmingly perceived as non-anonymous.
  • Confessions Drive Anonymous Content: A substantial portion of anonymous content (around 56.81%) consists of public confessions of guilt, shame, or embarrassment. This highlights a key use case for anonymous platforms as a cathartic outlet.
  • Linguistic Signatures of Anonymity: Anonymous content (whispers) is linguistically distinct from public posts (tweets). Whispers tend to be more personal, informal, and emotionally charged, expressing more negative emotions, wants, and needs.
  • Demographics Matter (Sometimes): While gender didn't show a strong overall effect, older users and those with higher education demonstrated a greater tendency towards anonymity sensitivity, particularly for sensitive content categories like relationships and NSFW.
  • Automated Detection is Feasible: The linguistic differences are significant enough to build machine learning models that can automatically distinguish between anonymous and non-anonymous content with reasonable accuracy.

Key Results

  • Anonymity Sensitivity Distribution: Approximately 16% of whispers have an AS score of 0.8 or higher, compared to only 1% of tweets. Conversely, 86.4% of tweets have an AS score below 0.3.
  • Content Categories: The top categories for anonymous content were Confessions (56.81%), Relationships (24.74%), and Meetup (17.82%). Categories like NSFW and LGBTQ showed particularly high anonymity sensitivity.
  • Linguistic Features: Key LIWC categories that distinguished whispers from tweets included "1st Person singular Pronoun," "Discrepancy" (expressing wants/needs), and "Sexual."
  • Classifier Performance: An automated system trained on LIWC features achieved an accuracy of 73.12% and an F1 score of 72.88% in distinguishing between anonymous and non-anonymous content.

Limitations and Future Work

  • Inter-Annotator Agreement: The study noted a poor inter-annotator agreement (Fleiss Kappa of -0.015), suggesting that anonymity sensitivity is highly subjective and varies significantly between individuals. This highlights the challenge of creating a universal measure.
  • Demographic Representation: While efforts were made to collect demographic data from AMT workers, the sample might not perfectly represent the broader internet population, potentially introducing sampling bias.
  • Future Work: The authors suggest investigating anonymity-preserving system designs that offer different levels of anonymity protection tailored to specific content sensitivities. They also propose developing automated anonymity advisors that alert users before they post highly sensitive messages. Further research could explore the impact of anonymity on user behavior in more depth and refine automated detection systems.

Summary updated

Abstract

Recently, there has been a significant increase in the popularity of anonymous social media sites like Whisper and Secret. Unlike traditional social media sites like Facebook and Twitter, posts on anonymous social media sites are not associated with well defined user identities or profiles. In this study, our goals are two-fold: (i) to understand the nature (sensitivity, types) of content posted on anonymous social media sites and (ii) to investigate the differences between content posted on anonymous and non-anonymous social me- dia sites like Twitter. To this end, we gather and analyze ex- tensive content traces from Whisper (anonymous) and Twitter (non-anonymous) social media sites. We introduce the notion of anonymity sensitivity of a social media post, which captures the extent to which users think the post should be anonymous. We also propose a human annotator based methodology to measure the same for Whisper and Twitter posts. Our analysis reveals that anonymity sensitivity of most whispers (unlike tweets) is not binary. Instead, most whispers exhibit many shades or different levels of anonymity. We also find that the linguistic differences between whispers and tweets are so significant that we could train automated classifiers to distinguish between them with reasonable accuracy. Our findings shed light on human behavior in anonymous media systems that lack the notion of an identity and they have important implications for the future designs of such systems.
BibTeX
@inproceedings{correa2021many,
  author = {Denzil Correa and Leandro Silva and Mainack Mondal and Fabrí­cio Benevenuto and Krishna P. Gummadi},
  title = {{The Many Shades of Anonymity: Characterizing Anonymous Social Media Content}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v9i1.14635},
}
Figure