Authors: Emilio Ferrara, Onur Varol, Clayton Davis,
Filippo Menczer, and Alessandro Flammini
Publication, Year: Science, 2016
The Rise of Social BotsIntroEngineered Social TamperingThe Problem with Malicious Social BotsThe Bot EffectAct Like a Human, Think Like a BotBehavior of Sophisticated BotsTaxonomy of Social BotsGraph-Based Social Bot DetectionCrowdsourcing Bot DetectionFeature-Based Social Bot DetectionCombining Multiple ApproachesMaster of Puppets
It's all Joseph Weizenbaum's fault (not really)
However, his creation of the software ELIZA, which utilized Rogerian psychotherapy methods to communicate with people, was one of the first social bots ever created.
Two types of social bots
Benign / Innocuous / Helpful: These types of bots include automatic email responders, automated response systems employed by customer service devisions of many companies, simple news feeds (i.e. RSS feed)
Malicious Entities: Designed specifically with the purpose to harm — these bots often mislead, exploit and manipulate discourse on social media with rumors, spam, malware, misinformation, slander and even by simply adding noise to the conversation.
These types of bots can cause real damage to society.
Misinformation is not new, however, the ability for bots to give content a false sense of popularity/importance can exert an influence on how individuals receive this information — a cognitive bias for which we have not yet developed tolerance to.
Highly sophisticated bots are also highly valuable because it may be impossible even for a human to distinguish the difference between a real account and a bot account
There is also evidence that social bots on Twitter can significantly impact the stability of markets
There is also evidence that bots can harm in more subtle ways
Social media users might be vulnerable to a social bot designed to get ask for personal information from people (phone numbers, addresses, etc)
Bots could hinder social movements by creating an inflated sense of popularity for a contrarian position
Can alter the perceived social influence of people for commercial or political purposes
A study has also show that emotions are contagious on social media and bots could be orchestrated for nefarious purposes and unpredictable public response — at least in theory
The boundary between human-like and bot-like behavior is now fuzzier.
Can search the internet for information and content to fill their profiles
Post collected material at predetermined times, emulating posting patterns that match typical posting rates for humans, and the overall temporal signature of the human circadian rhythm
Create/engage in conversations by posting in comments or on social platforms
Attempt to gain influence by acquiring followers
Can infiltrate popular conversations, generating topically relevant and even interesting content by searching for keywords within the discussion and returning relevant and "fresh" content for users within that conversation
Some bots can even tamper with individual identities
Even more advanced mechanisms can be employed; some social bots are able to “clone” the behavior of legitimate users, by interacting with their friends and posting topically coherent content with similar temporal patterns.
The field of detecting social bots was relatively new at the time (it still is, and is also ever-evolving) so they propose a simply taxonomy of the classes of bot detection methods…
Sometimes a hard categorization of a detection strategy into one of these three categories is difficult, since some exhibit mixed elements...
The challenge of identifying social bots has been framed in many ways. The Facebook Immune System represents one.
The above image was taken from here not the review paper which these notes focus on.
Proposed strategies to detect sybil accounts (social bots controlled by an adversary) rely on the structure of an accounts social graph (i.e. network of connections)
Unfortunately, this method is not very practical because the central premise is that people would not connect with an account of someone that they do not know. Yet...
Thus, Alvesi et al (2013) suggests we employ a portfolio of techniques with the manual collection of legitimate accounts to aid the development of machine learning tools
Wang et al (2013) created an Online Social Turing Test platform to try and utilize human beings to identify social bots on Facebook and RenRen (a popular Chinese platform)
Unfortunately, there are three drawbacks to this method
Difficult for big companies
Difficult for small start-ups
Feature-based models typically utilize machine learning techniques to encode bot-like and human-like behaviors.
A few benefits of utilizing feature-based methods:
An example of a feature-based approach is Botometer (formerly Bot or Not?).
As social bots change over time to avoid detection, feature-based methods allow a flexible framework which can be updated and improved upon based on new feature patterns with observed data
Wang et al developed the Renren Sybil detector to implement a system which combines a multitude of the approaches described herein.
This approach has a "Sybil-until proven otherwise" has the benefit of potentially picking up new and as yet unseen bot-like patterns
Other methods which utilize a mixed methods approach are:
If social bots are the puppets, additional efforts will have to be directed at finding their “masters.”
Understanding who is behind malicious social bots will be a difficult problem to solve.
Studying platform vulnerabilities has begun as path of research as well
Unfortunately, many basic research questions remain open
A constant limitation of supervised machine learning methods is the need for new and relevant training data since bots can change much faster than scientifically rigorous research can be conducted...
Notes by Matthew R. DeVerna