Gray networks are semi-clandestine and have an organizational structure that is only partially known. They may have some secrecy to their membership, use aliases, have initiation rituals, engage in illicit behavior, and so on, but they aren’t classified as strictly a criminal enterprise like a traditional “dark” network.
Gray networks are somewhat understudied in the social networks and dark networks literature.
In this paper, I use publicly available social media trace data from Venmo and Facebook to explore the structure of the Proud Boys as a “gray network”.
The Proud Boys Venmo network as it grew between October 2016 and April 2019
My main research questions were around whether we can predict the leadership of the group solely from the features of the network, and what can we learn about the way the real world group is organized by looking at the online social network and comparing that to the actual leaders of the group, for example as revealed by their un-redacted Bylaws document.
Proud Boys financial network on Venmo, as of April 2019. Black nodes represent 5 of the 8 “Elders” who were also using Venmo.
Findings of this paper include:
Proud Boys social network data confirms their geographically-oriented structure. In addition, the network predicts that 8 leaders (“Elders”) were probably chosen to represent 8 main geographical areas.
Social media trace data reveals that real-world events (e.g. TexFest, WestFest, various rallies) are a main driver of membership, payments, and leadership.
Proud Boys social network data reveals both bridges and hubs in the structure, as well as small world features. The presence of these different structures can indicate their preferences for how Proud Boys chose to address the communication-versus-security tradeoff characteristic of clandestine networks.
Two of the cells appear to operate largely independently of the larger nationwide structure.
Although only 5 of 8 Elders appeared in the financial network, 4 of these 5 were predictable as leaders based on network metrics alone.
1 of the 5 Venmo Elders was not able to be predicted from the social media trace data.
4 additional “leader” nodes (2 hubs, 2 bridges) occupy influential positions even though they are not known to be Elders.
This paper has been accepted for presentation and publication at the 11th International Conference on Social Informatics, to be held in Doha, Qatar November 18-21, 2019. (double-blind refereed, 24% accept rate)
Donald Trump at his Social Media Summit in July, 2019
In this policy piece for Brookings Institution, I lay out the problems with content moderation of white supremacist, extremist content as it exists now. I then suggest several avenues for alternate approaches. First among these is an admission that this problem is not easy, and will require a subsequent serious investment in real, deep expertise.
Here is a podcast I recorded with Al Letson from Reveal about the dark side of “Alt-Tech” – the suite of Internet infrastructure tools being developed for a censorship-free Internet but which are already being used by extremists to spread their hatred.
I’m in the first 15 minutes, and followed by a story about alt-right comics and some good words from a pastor working on the front lines of extremism.
After the horrible events in Christchurch, New Zealand last month, I spent some time tracking the terrorist’s manifesto and video as they traveled through the web. I wrote this article describing the technical aspects of how to track a file and why it is difficult, and some technologies -such as IPFS- that extremist groups will be using in the future to make removal of a file even more challenging.
Side note: I originally wrote the article for The Conversation, but since everything there is CC-licensed, the story was picked up and re-run by International Business Times, Raw Story, Alternet, International Policy Digest, and so on. I suppose that is an interesting meta-story about how my own article moved through the web.
I wrote this article to describe how 202 anti-Muslim groups (US-based) on Facebook during the period June 2017 – March 2018 use text and images to describe their cause.
The TLDR is that violence and “enemy” attitudes are the predominant text and visual indicators in these groups.
I wrote this little paper after being frustrated at the lack of data about precisely how many women were in far-right groups these days, especially online. I was hearing estimates ranging between 7% and 56%, and some of these didn’t take social media into account at all, but were just back-of-the-envelope guesstimates based on event attendance or interviews with group members.
In this work, I use a very large collection of data I collected about far-right extremist group members from Facebook’s API during the period June 20, 2017 – March 31, 2018. I then used two “genderizer” software packages to infer the gender of these 700,000 extremist group members. I then divide those into ten different ideologies and look for evidence of women’s auxiliaries, sometimes mockingly called “wheat fields”.
I find that wheat fields DO exist in five of the ten ideologies. The tall spikes at the left side of each graph below represent groups with a super-majority of women, and which are designed to be women-oriented groups.
I find that women’s leadership rates in the Facebook groups differ between ideologies. One smaller ideology, Neo-Nazi, tends to use women in leadership roles at a higher-than-expected rate. White nationalist and Proud Boys, both ideologies with wheat fields for women (a.k.a. women’s auxiliaries) don’t tend to have women in leadership roles in groups as a whole.
Finally, just like in the 1920s with the WKKK (Women’s KKK) and the Ladies’ Memorial Associations (LMAs) and United Daughters of the Confederacy (UDC) I find a systematic marginalization and oppression of women on the one hand comes into conflict with a practical need to leverage women’s networks and organizational abilities on the other hand.
This research was also recently covered in Buzzfeed News.
Below is a network diagram showing some of the extremist groups and ideologies in my data set, and how they overlap in membership.
Network of Anti-Muslim & other extremist groups on Facebook, with 10+ members in common (click to enlarge)
Two of the key anti-Muslim groups in this network – each scoring very high on betweenness centrality measures – are Infidel Brotherhood International and Stop the Islamization of America. Each of their ego graphs are shown below:
Anti-Muslim groups attract the same audiences as other extremist ideologies, including secessionist neo-Confederates, militant anti-government conspiracy theorists, and racist white nationalists. In addition, groups like IBI and SOIA can serve as a convenient lingua franca: their brand of hate is a common denominator that ties extremists of disparate ideologies together.
I’ve updated the Facebook co-membership graphs (see original post) for my upcoming talk at the International Conference on Computational Social Science (IC2S2) to be held at Northwestern University in July. (extended abstract – PDF)
This talk will include data through the end of March, 2018.
Once again, larger nodes = more people. Closer placement between nodes on the graph mean more folks in common.
What do we learn? There are some ideologies that are woven much more naturally into the fabric of a “united” far-right, as opposed to other ideologies, which will be harder to integrate.
Upcoming work will look at groups with nativist ideologies, including anti-Muslim, anti-Immigrant, and how those correspond to Anti-Government/Patriot/Militia and White Nationalist beliefs.
As Wired recently reported, I have built a large dataset of far-right extremist groups in order to learn about their movement and figure out how they operate so that we can keep our communities safe.
Where does the data come from?
Much of my data comes from public, online sources. For example, despite Facebook claiming to be “no place for hate”, I have found that the Facebook social network is actually a very rich source of data about hate groups and the extreme far-right. For 500,000 members of 1,336 different far-right groups and events, Facebook is the perfect place for recruitment and community-building.
No place for hate?
In this blog post, I describe how I divided these 1,336 groups into 11 different far-right extremist ideologies (following the SPLC’s and ADL’s descriptions of these beliefs), and performed some social network data mining to answer questions about group co-membership.
What can we learn from this data?
Since the far-right was attempting to “unite” itself by holding events such as the August 12 Charlottesville rally (recall that this event was literally called “Unite the Right”), I wondered: Do groups from different ideologies share members in common? Which groups have the most members in common?
What are the different ideologies represented?
I originally had 9 different far-right ideologies in the classification system, and have since expanded to 11, as shown in the table below.
To classify each group or event, I evaluated the text and emojis used the group name and description, as well as text and symbols found in the group cover photo, and the public content if viewable (discussion comments, photos, and so on.)
Primarily, I use Facebook’s own recommendation system for finding new groups (“Suggested Groups”). I also use their search API for groups with keywords and concepts related to a particular ideology.
Special note since some right-wing conspiracy theory web sites are misreporting this fact: Keywords alone are NEVER sufficient to classify groups or to select them for inclusion. With Google now indexing over 620 million Facebook groups, and only 1,336 in my database, obviously I am not just randomly including every group that matches a keyword! That is a ridiculous assertion. For example: “rebel” is a very generic word, and there are many thousands of groups with that word in the title. As explained above, I follow the SPLC’s and ADL’s descriptions, and I manually evaluate every single group for a match. If a group doesn’t clearly qualify (by name, photo, content, and/or description), that group CANNOT be included.
Ideology
Count of groups
Concepts related to this ideology
Neo-Confederate**
347
Confederate, League of the South, rebel, confederate flag, secession, dixie, CSA (full description)
UPDATE: After getting feedback about the way I had classified Christian Identity and The Creativity Movements, I decided to collapse the category, moving Christian Identity into Anti-Semitic as SPLC recommends, and likewise, The Creativity Movement is moved into the Neo-Nazi parent category. This makes a lot more sense. Thank you, readers. I guess I’m back to 11 categories again! Numbers have been updated in the table above.
Social Network Analysis
Below is a diagram (created in Gephi using Fruchterman-Reingold layout) showing co-membership between groups from five of the ideologies: Alt-Right, Neo Confederate, White Nationalist, Anti-Government, and Neo-Nazi. Lines (“edges”) between groups (“nodes”) indicate that the two groups share at least 10 members in common. No group is included if it does not share at least 10 members with another group. Larger nodes mean the group has more members in it.
click to expand
In the center of the diagram, in purple, is the Unite the Right (UtR) event. Below is a close-up view of the groups and events that had the highest co-membership with UtR. Here I have highlighted the UtR node, and its connected nodes show up with their color (nodes that are disconnected from UtR in this diagram are “greyed out”).
click to expand
As the diagram shows, UtR not only had a very high number of nodes it shared members with, but the nodes it shared with were from a wide variety of groups from all the other ideologies: Neo-Confederate (red) groups, militias and Oath Keepers (blue), White Nationalist groups (green), Alt-Right groups (yellow), and even Neo-Nazi groups (black).
Which other groups are similarly well-connected? There are a few, but none cast as diverse a net as UtR. The figure below shows some of the largest nodes, all clustered in the busy center of the diagram. We seee a large neo-Confederate group, a very large militia group, a /pol/ (4chan) Alt-Right group, an “anti-SJW” (against “social justice warriors”) group, a “white culture” forum claiming to be the largest on the Internet, an Odinist/Folkish style group, and a forum for National Socialist ideas.
click to expand
More about classifying groups by primary ideology
The problem of multiple ideologies. Many far-right extremist groups could easily be classified into more than one ideology. Most of these groups are anti-Muslim, misogynist, anti-Semitic, and so on. Therefore, to come up with a group’s primary ideology, I will typically rank the name and description as of higher importance than the cover photo when trying to make the call. Still, there are gray areas. I once saw a group that listed 5 different ethnicities it disliked, used a militia symbol on the cover photo, and had a generic, non-descriptive name. Classifying that group was a challenge.
To handle the multiple-ideology problem early on in this project, I originally developed a tagging system where I could tag groups with multiple keywords. In that system, a group could be classified as “militia / anti-Muslim / folkish” or “neo-Confederate / militia / anti-immigration”. This system quickly became untenable, however. It was just too difficult to generate meaningful graphs when every group could be in multiple categories at once. At this point, I think such a tagging system should be built on top of – or in addition to – a primary classification system.
The problem of changing ideologies. Each time a group changes its name, description, or cover photo, it could change the perception of what the group’s primary ideology is. I have run into a few groups that went from neo-Confederate to militias, from anti-Muslim to anti-Government, from anti-Semitic to Alt-Right, etc.
Combining ideologies. Depending on the analysis type I want to do, I might sometimes combine some smaller groups into a larger group. For example, if I’m interested in studying “whiteness” groups, I might combine the large white nationalist category with the smaller Christian Identity and racist skinhead categories. However, if I’m specifically studying Christian Identity groups as a niche community, I can keep these ideologies separated.
Over-representation of certain ideologies. Since I live in the southern portion of the United States, I tend to find more Neo-Confederate groups than certain other types of groups that are more prevalent in other areas of the country.
In future blog posts, I’ll show some of my other data sources, as well as other ways to analyze the Facebook data.
