Global Security and Intelligence Studies Volume 5, Number 1, Spring / Summer 2020 | Page 64
Global Security and Intelligence Studies
often, which might be less appealing to
IFO actors because there is no schedule
for users and IFO actors to follow.
In general, the highest percentages of
anomalous users were from channels
with the largest number of total users,
possibly because a larger audience is
a better target for IFOs and/or spam
posting.
The study highlights the need for
bot-hunting artificial intelligence (AI),
as bots are becoming increasingly complex
as technological advancements
are made. For example, an in-depth
IFO-detection study must utilize more
than just comment count and comment
speed to identify bots, as clever
IFO actors could adjust their bots to
post no more or no faster than some
pre-determined limit (say, the stream’s
current mean or median posting speed
or count). IFO actors could also use AI
to generate comments for their bots,
rather than have bots execute comments
from a pre-established comment
bank. Finally, if an IFO actor develops a
bot that posts on a completely random
schedule, dynamically generates content
analyzed from ongoing streaming
audio, visuals, and comments, and actively
responds to users, a human analyst
will be virtually incapable of identifying
the bot. Overall, the complexity
of future bots needs to be met with the
complexity of AI—AI will be needed
to recognize advanced bot algorithms
(Manheim and Kaplan 2019).
The authors acknowledge two
major limitations with this study. First,
it is difficult to determine whether or
not a Twitch user is a bot—humans do
not possess the ability to distinguish
bots from humans except in blatant
cases (for example, if a bot posts the
same or similar messages at fixed intervals).
Second, the criteria may have
excluded possible slow-posting bots.
Future researchers could develop (or
incorporate existing) machine learning
and sentiment analysis programs
to further refine bot search criteria on
Twitch. Additionally, researchers could
develop a bot-detection metric or criteria
checklist to allow for manual or
automated assessment of users, rather
than a subjective look over. Finally,
researchers should search for countermeasures
that actors employ to protect
their bots from discovery.
Conclusion
The purpose of this study’s was
to develop a data-mining prototype,
rather than develop a
reliable and effective bot-identification
program. We did not seek to prove the
existence of IFOs on Twitch, but rather
show it is possible to identify them if
they do exist, and encourage future researchers
to use some of our methods
to narrow their bot searches. We maintain
that our research provides future
Twitch IFO- and bot-hunters a better
starting point for discovering IFOs and
bots.
As the internet audience
grows, the potential for IFO development
and execution grows. State
and non-state actors know the value
of IFOs during peacetime and wartime.
Twitch is only one vulnerable
platform. Online multiplayer games
50