Global Security and Intelligence Studies Volume 5, Number 1, Spring / Summer 2020 | Page 61

Discovering Influence Operations on Twitch.tv: A Preliminary Coding Framework
Table 1: January 29, 2020 Chat Log Download
Channel
Stream Type
Anomalous
Streams
Total
Streams
Anomalous
Users
Total
Users
%
Anom.
Users
bastiat Political 6 10 83 2610 3.18
Bernie_Sanders Political 0 10 0 3117 0
DemocracyLive Political 6 10 28 1105 2.53
DonaldTrump Political 0 5 0 9471 0
hasanabi Political 7 10 2101 37994 5.53
skynews Political 1 3 18 1294 1.39
touringnews Political 1 10 5 670 0.75
washingtonpost Political 9 10 748 14203 5.27
hutch
JakenbakeLIVE
Alinity
badbunny
Political/
gaming
Political/
live blogging
Live blogging/
gaming
Live blogging/
gaming
3 10 179 4378 4.09
10 10 2412 40269 5.99
10 10 486 13588 3.58
7 10 227 6952 3.27
ninja Gaming 6 10 658 37628 1.75
riotgames Gaming 3 10 387 13568 2.85
Total 69 128 7332 186847
6. IFO actors target conversational
streams with automated comments.
If automated comments are engaged
with, human actors can take
over for manual commenting.
7. IFO actors target rapid-posting
streams with short messages in a
quantity-over-quality approach (e.g.
spamming the hashtag “#FREE
TIBET” in chat).
8. In conversational and rapid-posting
streams, IFO actors post more than
the stream’s norm, as they are trying
to make their comments stand
out against the rest of the chat. By
posting more frequently and/or in
higher volumes, IFO actors’ comments
are identifiable via statistical
techniques.
We selected 14 Twitch.tv channels
for analysis, and included political
or apolitical content creators (see
Table 1). We handpicked channels to
confirm proof of concept, rather than
to execute a completely unbiased study.
We chose popular channels because
IFO actors likely want to target many
users at once. We expected that political
channels would have a greater bot
47