1/n
Alternative headline: "Deceptive use of deep learning to permaban Twitter users you don't like." As someone who works on both the spread of disinformation and privacy, I can say a lot about this. Let's go.
elpais.com/tecnologia/202…
Alternative headline: "Deceptive use of deep learning to permaban Twitter users you don't like." As someone who works on both the spread of disinformation and privacy, I can say a lot about this. Let's go.
elpais.com/tecnologia/202…
2/n
Let's get this out of the way: disinformation is a real problem that needs to be addressed & disinformation about covid-19 is amplifying the crisis. We should be working on ways to dampen disinformation and promote quality journalism.
Let's get this out of the way: disinformation is a real problem that needs to be addressed & disinformation about covid-19 is amplifying the crisis. We should be working on ways to dampen disinformation and promote quality journalism.
3/n This article argues that one method to avoid the spread of disinformation is to record the keystroke patterns so that when you catch someone spreading disinformation you can recognize them and ban them even if they create a new account to get around your ban.
4/n Typing works like a digital fingerprint to identify you. There are other fingerprints (writing style,browser fingerprinting), but unlike actual fingerprints, these methods often have high false positive rates, more so when you scale to 1.73 billion or so users.
5/n More important than what the false positives rates are, is what a FP looks like and which users it affects the most. In this case (similarly with writing style), false positives will be users who type like someone who shared a piece of fake news about covid-19.
6/n I don't know of any specific research on typing, although I suspect that, like writing style, it correlates to native language (and other traits). A FP here means censoring someone because they type similar to someone who shares disinfo (and maybe from the same demographic!)
7/n While we're close to bias, it is likely the case that false positives are not distributed evenly among classes. That is, particular demographics (women, the elderly) will be disproportionately flagged.
8/n Pretend for a moment that you're a troll who, for whatever reason (🤑), is spreading fake news about covid-19. You get reported and banned. That's ok, you create a new account and start from scratch. Oh no, the social network finds you by analyzing how you type your post!
9/n The defense here is so simple that you already know what you're going to do next: type out your post in a notepad and copy-paste. All positive samples are soon either FPs or people who are not maliciously sharing fake news, but who are victims of the fake news machine.
11/n Let's talk about those people real quick. Sure that one cousin who shared the plandemic video and hates bill gates for some reason is really annoying. Is the solution to ban them from Facebook?
In the mean time, you're still collecting behavioral data from innocent users...
In the mean time, you're still collecting behavioral data from innocent users...
12/n A lot of people are sharing conspiracy theories right now because they're scared and they're accepting things as true that they wouldn't otherwise. A permaban from social media might help in the short term, but it's a punishment beyond the crime. #quitfacebook (#quittwitter)
13/n On privacy: Boy oh boy the privacy paragraph in this article. First, for this to work you need to collect every post of every user: even those who are not malicious! Who collects the data? Where is it stored? When is the model trained? Who has access to it?
14/n "If it falls into the wrong hands." What are the right hands for this? Who has the right to decide who has a voice on social media? Does banning = censoring now that public discourse is almost completely online?
15/n If we plan all protests, marches, & boycotts on Facebook, lack of an account removes you from the conversation. The solution that I'll advocate for of course is that we need to organize off of Facebook. Another soap box for another day. Back to privacy. Oh GDPR!
16/n It's a fallacy that GDPR protects all privacy. GDPR regulates the use of data. As researchers, we cannot think: "how much can we do and still be compliant?" but instead "how can we do our task in the most privacy preserving way?"
17/n The potential for misuse here is immense. Sure, a lot of research is wide open to misuse (mine included!) but having an infrastructure in place to ban users who make new accounts means that you forfeit that anyone can change. And this is what I find most frightening.
18/n Final soapbox: All CS conference/journals should require an ethics section that discusses the potential negative impacts of the work, including societal impacts of correct use, misuse, and possible side effects. this paper has none
This was my first Twitter thread! How fun.
This was my first Twitter thread! How fun.
