@gianluca_string is up at #enigma2021 speaking about "COMPUTATIONAL METHODS TO UNDERSTAND AND MITIGATE ONLINE AGGRESSION"
[Make sure to catch the talks when they're posted -- my hands can't keep up with these speakers and the talks are 🔥]
usenix.org/conference/eni…
[Make sure to catch the talks when they're posted -- my hands can't keep up with these speakers and the talks are 🔥]
usenix.org/conference/eni…
Content note: this talk is about online abuse as some of the content may be upsetting
Got pulled into this after a screenshot of a class assignment sending folks to post on 4chan to post about race/gender/etc issues got posted on 4chan without the email address... so the 4chan folks thought it was @gianluca_string. It wasn't, but they doxxed and harassed anyway
This coordinated attack is a "raid". If even white, male profs are getting these, imagine who else gets them worse
~2-5/day
~2-5/day
How can computer security people help?
* we study coordinated malicious activity by code
* this is coodinated malicious activity by people
Can we use techniques from one on the other?
* we study coordinated malicious activity by code
* this is coodinated malicious activity by people
Can we use techniques from one on the other?
What is 4chan?
[If you don't know, that may make you the winner. Especially if you don't know about 8chan.]
* public, anonymous image board where content ages out
* released a dataset with all the posts
[If you don't know, that may make you the winner. Especially if you don't know about 8chan.]
* public, anonymous image board where content ages out
* released a dataset with all the posts
Threat model
* someone calls for attack
* organization thread
* delivery across many platforms
* harm
* someone calls for attack
* organization thread
* delivery across many platforms
* harm
Focused on YouTube because it's the most popular platform linked from 4chan
Identifying raided videos on YT -- spike in hate comments
[note: this is using some kind of API to identify hate speech, which has some real limitations, especially on certain dialects]
[note: this is using some kind of API to identify hate speech, which has some real limitations, especially on certain dialects]
What to do about raids?
* too many videos, too many comments, can't look at them all
* developed a risk score which can identify with AUC .94 whether a video will be raided at upload time
* too many videos, too many comments, can't look at them all
* developed a risk score which can identify with AUC .94 whether a video will be raided at upload time
Latest threat: zoombombing
* attacker posts link to meeting room
* people join to disrupt
* attacker posts link to meeting room
* people join to disrupt
Measuring zoombombing
* pulled out links from 4chan and twitter
* labeled posts
* analysed
* pulled out links from 4chan and twitter
* labeled posts
* analysed
74% of the calls for zoombombing target classes, 70% by students in the class [for the lulz]
93% target meetings happening in real time
93% target meetings happening in real time
We need to rethink our defense mechanisms
* passwords don't work (because the insider can pass it on)
* waiting rooms don't help (the insider attackers supply names of real students to spoof)
Instead use one-time unique links [note that Zoom has this in some modes]
* passwords don't work (because the insider can pass it on)
* waiting rooms don't help (the insider attackers supply names of real students to spoof)
Instead use one-time unique links [note that Zoom has this in some modes]
[end of talk]
• • •
Missing some Tweet in this thread? You can try to
force a refresh
