Hacker News Reader

Hacker News

Comment by ngngngng | Hacker News Reader

Parent Full thread

ngngngng

> captures 99% of child exploitation and terrorism-related content

Credit where it's due, good on them building such effective systems to catch this content.

View on HN

apropos: https://i.imgur.com/djxjXbj.jpg

* https://en.wikipedia.org/wiki/Four_Horsemen_of_the_Infocalyp...

Image filtering is now trivial even using simple transfer learning, leading to >99% accuracy. It's just the training datasets might be super disgusting and question one's will to live when confronted with the savagery that is going on around the world.

Wasn't there some issue a while ago that the UK police were having trouble distinguishing pictures of sandy deserts from nude bodies?

How is it not a crime to possess the training set?

Should be simple enough to have the government/LEAs posses the training data and have a standard harness/workflow you integrate your training program to. Then you send your training peogram and get the resulting trained model files back.

AFAIK large companies have deals with the government and have special units that deal with that type of content (nobody survives longer than 12 months there).

You can get legal datasets for sick pr0n or warzone stuff you would like to keep off-platform that might have the same psychological effect on you.

I assume that's capturing 99% of the child exploitation that is captured, not that's estimated to exist on their platform. It might be good for them, or bad for the government, or par for the course.

But they just made that number up.

more or less agreed, would def like to see how they ended up at 99%

Right. Assuming that they know that they're missing 1% of cases insinuates that they know enough about the 1% they're missing to quantify it.

Isn't that right based on the statement though? The implication is 99% of the identified cases on Facebook are caught by Facebook's systems, the 1% not caught by Facebook's systems are tracked by other means. 100% of identified cases, not 100% of all cases.

Still very misleading if that is the case, what other means are there realistically to catch something if not by FBs systems, on FB? Self-reporting in the groups? Is that counted towards the FBs 99% or the 1%?

Reported directly to the police or other authorities? It's probably quite realistic in those scenarios for something that Facebook's systems have missed for it to be found by someone eventually and escalated.

There's a little ambiguity in there, but I wouldn't go so far as to say it's "misleading".

Is it made up? I thought they meant "Of content identified on Facebook, 99% was identified by Facebook's own systems". Not that Facebook captures 99% of terrorists. Just that it does the bulk of capturing terrorists _on Facebook_.

99% of all uses of the number “99%” are made up.

And what's the false positive %. How do you even define a genuine hit for "terrorism-related". Depending on how broad you go that can mean a lot of things, including very innocent stuff. The problem with terrorism is that everyday things that are legal on their own when used as intended can be used to commit those acts.

Reply on news.ycombinator.com