News for hackers. Stuff that matters.

A comment on Inside Pornhub. Parent comment

Re:Inside Pornhub by tshaddox on Wednesday February 23, 2022 @05:34PM (#30447505)
I don't follow the argument that the false positive rate necessarily goes up if you notice some successful attempts to defeat your system and update your system to detect some attempts. Is this some statistical law you're referencing, or do you just mean that people tend to do naive updates to these detection systems to make them more sensitive (not necessarily "more sophisticated")?
[ Parent | on HN ]
Re:Inside Pornhub by mattnewton on Wednesday February 23, 2022 @08:05PM (#30449043)
More to the point, I don't buy that a higher false positive rate isn't worth it in this application. From a social perspective, the inconvenience of the false positive rate seems so small as to be impossible to measure on a scale weighed against the harms of uploading revenge porn.

I know the companies are optimizing for other metrics where the inconvenience might be measurable but I think that's a failure of regulation then. The platform could be made liable like we do with copyright material, for example, and then like YouTube they would be incentived to accept a higher false positive rate and automatically filter repeat offenders out.

[ Parent | on HN ]
Re:Inside Pornhub by armchairhacker on Wednesday February 23, 2022 @07:33PM (#30448762)
YouTube has definitely been flagging more false positives over the years, as pirates have gotten smarter. When you broaden the algorithm to flag more bad content, you also end up flagging more good content unless you narrow it in some other way.

But the issue with YouTube etc. isn't flagging innocent people, it's flagging innocent people and then rejecting their appeals. It should be pretty easy for PornHub etc. to get videos "verified" by getting explicit consent from all parties filmed, so that they cannot be flagged and removed from the system.

In fact, I really don't see any reason against a law requiring all porn sites to require verification for all videos (w/ retroactive revoke on the hosts). Nobody's being censored, and if you agree to be in a porn video you kind of waive your right to anonymity.

[ Parent | on HN ]
Re:Inside Pornhub by AnthonyMouse on Wednesday February 23, 2022 @10:50PM (#30450007)
> Nobody's being censored, and if you agree to be in a porn video you kind of waive your right to anonymity.

Then the host gets hacked and your government ID gets published as a creator of amateur pornography. The risk of that is de facto censorship by chilling effect.

[ Parent | on HN ]
Re:Inside Pornhub by pjc50 on Thursday February 24, 2022 @05:11AM (#30453017)
The thing is, this was technically required in the US under USC 2257 for the past two decades: https://www.law.cornell.edu/uscode/text/18/2257

Although I think that says "produces" rather than "publishes" or "distributes", so pornhub themselves aren't liable but the boyfriend "producing" the film has committed an offence. Not that there's any will to enforce this.

Closer to home, Scotland explicitly banned revenge porn, and has actually achieved some prosecutions.

[ Parent | on HN ]
Re:Inside Pornhub by BugsJustFindMe on Wednesday February 23, 2022 @07:49PM (#30448908)
> Is this some statistical law you're referencing

There is a natural law in pattern recognition governing the split between sensitivity and specificity in scenarios with uncertainty, though I don't know if it has a special name. Catching more edge cases in a poorly-defined group requires allowing the inclusion of more things that look a lot like edge cases, but cases on the edge of a group are by definition similar to things outside of the group. The more you want to eliminate false negatives the more likely you are to produce false positives.

[ Parent | on HN ]
Re:Inside Pornhub by sdenton4 on Wednesday February 23, 2022 @11:22PM (#30450234)
In binary classification, this is basically a precision/recall trade-off. More precision (ie, fewer false positives) necessarily means less recall (ie, more false negatives). Classifier statistics like AUC basically tell you how good the classifier is in terms of how bad the trade-off is.

Adversarial situations - any time humans are trying to trick your classifier, in other words - will always make the trade-off worse.

[ Parent | on HN ]
Re:Inside Pornhub by Spooky23 on Wednesday February 23, 2022 @06:29PM (#30448158)
It’s just a convenient and lame excuse. Slippery slope stuff is the worst/laziest argument.
[ Parent | on HN ]