Comment by cortesoft
1 day ago
Yep, and this is a perfect example of a base rate fallacy situation... even if the scanner is 99.99% accurate, because an even higher percentage of photos are innocent, most matches the scanner will find will be false positives.
Funny you bring this up.
Back in the day when I was like 15 and DC++ was still a thing, I used to browse people's shared folders. One day I came across a file called "the paradox of false positive". It was a 1 pager that described how a machine which is 99.9% accurate at identifying terrorists would be completely useless due to this false positive base rate fallacy you're describing.
It really stuck with me throughout the years. It's kind o remarkable how even a 99.9% accurate heuristic is insufficient at scale.
Which begs the question: lets assume the intentions are pure (which we know they're not but lets be generous), what other options are there when 99.9% heuristic is not good enough? how do you design systems when they're guaranteed to fail as they scale up?
edit: and what do you know, I just saw this as I scrolled down on HN https://news.ycombinator.com/item?id=48816959
The system we got for this is called parenting.
And there is a saying where I grew up: you need a village to raise a kid, I feel like we lost track of that and feel the issues of that now.
Btw, von der leyen is trying to get stuff like this written down as laws since 2009, it got her the nickname Zensursula.
>Btw, von der leyen is trying to get stuff like this written down as laws since 2009, it got her the nickname Zensursula.
And Germans and Europeans looked at that and thought the best place for her is leading the EU?!
Remind me again how she got elected in that position?
Because it seems like the entire EU population knew her being infamous for that, except for the few elites who appointed her there via "democratic process" to the head of the EU.
5 replies →
> The system we got for this is called parenting
And it fails miserably.
9 replies →
The intuition I've built is that you can't talk about a false positive rate being high or low on its own - it's always relative to the actual occurrence rate of positives in the tested population. E.g. if there's a 1 in 10000 risk of a false positive, but real positives also are only 1 out of 10000 tested cases, then a positive case will have a 50/50 chance of being a false positive (because for every 10000 tests, you'll have on average one false positive and one real positive). So a false positive rate can only be said to be low if it's significantly lower than the real occurrence rate of positives.
The mentioned accuracy in the comment you are replying to already encapsulates the relation of true positives to false positives.
5 replies →
Because no one around me understood how to calculate false positive/false negative probabilities, I put together this calculator[1] in 2020.
[1]: https://www.covid2020.icu/false-positive-false-negative-simu...
I thought this was known as Bonferonni'a principle? Or am I getting mixed up?
Bonferonni correction is relevant when you calculate multiple p-values. Most statistical tests are used with a p-value threshold of 5% to reject the null-hypothesis. But because you are repeatedly testing, the probability for false positives increases and that is why you need to decrease the threshold and make it harder, to obtain a p-value below that threshold to declare a significant result.
You typically use the Bonferroni correction when making general statements about a statistical relationship. You wouldn't use it for checking if a particular image shows illegal content. If you kept testing with your image classifier, your significance threshold would need to be continuously lowered and you would asymptotically reach zero.
Relevant XKCD: 882
Google have already caused significant hardship to a father for such kinds of photos. What's particularly galling is how they've continued to maintain they were in the right, despite the police saying no crime had been committed.
https://www.koffellaw.com/blog/google-ai-technology-flags-da...
Of course google and every other big-tech platform is gonna insta-wipe every account containing detected nudes of children, regardless if you're the parent.
The corporate liability of such content being found on their cloud is so insanely nuclear, that they're not gonna wait and ask you "hey are those nudes your own kids or are you a pedo?" before they wipe the account with all pics off their servers.
And yet the will badger you endlessly to the point their photos app is near unusable to turn on auto sync which slurps up every photo and makes it very awkward to then delete them after. To me, this makes Google a liable party even if real CSAM is stored.
> even if the scanner is 99.99% accurate, because an even higher percentage of photos are innocent, most matches the scanner will find will be false positives.
If the scanner is 99.99% accurate, then most classifications will be correct.
If you scan 1,000,000 pictures (with let's say 10 CSAM), you'll have 100 false positives and 10 true positives, giving you like only 10% correct results
Ah, sorry, I misread what the OP meant by "matches" - thought they were referring to all classifier outputs, while they specifically meant the positives. I changed my original comment to better reflect what I meant, even though that makes it a bit of a non-sequitur now.
https://en.wikipedia.org/wiki/Base_rate_fallacy
How many child abusers do you think there are out there?
Even if 10% of population were actively criminal pedos (which is waaaay too high), its pretty safe to assume that the majority of even their online footprint would be ordinary images/messages.
So a quota of 0.1% or even less material being detectably criminal sounds realistic (probably not much less, though).