Coincidentally I was trying this out yesterday as a potential option for automated porn filtering for the imageboard that I administrate. It is very lightweight which is good but the performance is not great. Some weird false positives and it does a bad job with flat chested women which is very concerning to me because it leads me to think it won’t be effective against CSAM.
I guess CSAM detection will remain hard as the training data raises so many ethical concerns.
I wonder if there's a way in which we can forward-hash sensitive material like CSAM and train a classifier that only consumes the 'hashed' version of the material for training and detection.
Very naive approach but wouldn't training a model on a few specific buckets like "someone is naked", "image is explicit" and "there is a child" in the picture, would do the job without having to train explicitely on CSAM? If the model returns both a high probability of "there is a child" and a high probability of any other bucket than this image is classified as CSAM. It'd be high recall,low precision but one would be on a rather safe side.
This is already done, there's hashed indexes of known CSAM and you can apply vector steering to the latent space to shift detections to trigger on CSAM material. You can also use two detectors for age estimation and a separate one for nudity and use them in combination.
it is amusing that a model from a country that has banned porn would ever be developed. Wouldn't there be hurdles for them to access "data" at every stage?
(I know they must have used a VPN for training but it does seem like an added hoop. OFC the UAE's block is also not particularly thorough.)
You'd assume such models would do better in a laissez-faire nsfw environment than in a walled garden of relatively-less sin.
It's a story as old as time so I'm frankly less amused by it. The old school censors always saw every "data" that was taboo under the sun prior to censoring it for public consumption.
The number of downloads really doesn't seem like a good way to measure the quality of a bit of work. One download could be deployed on a bit website and classify billions of users images.
So in a bit more than a year, 10% of the planet has downloaded a NSFW classifier?
I mean: in the end there are only 8 billion people on earth, which projects do, all added up, need to download one NSFW classifier 600 million times a year?
Something something something makes me think there are inefficiencies at play here.
Depends, if every time a project redeploys it redownloads all external resources from their respective CDNs, do you consider that to be inefficient? What if it’s part of a service that runs on thousands of instances to keep up with the volume of user generate content?
I wonder if there's a way in which we can forward-hash sensitive material like CSAM and train a classifier that only consumes the 'hashed' version of the material for training and detection.
The jersetters are basically exempted from the strict laws because they bring in money.
Alcohol is illegal there, homosexuality is, even unmarried sex is.
I certainly will never visit there even if they choose not to apply those laws to me as a visitor.
(I know they must have used a VPN for training but it does seem like an added hoop. OFC the UAE's block is also not particularly thorough.)
You'd assume such models would do better in a laissez-faire nsfw environment than in a walled garden of relatively-less sin.
The number of downloads really doesn't seem like a good way to measure the quality of a bit of work. One download could be deployed on a bit website and classify billions of users images.
So in a bit more than a year, 10% of the planet has downloaded a NSFW classifier?
I mean: in the end there are only 8 billion people on earth, which projects do, all added up, need to download one NSFW classifier 600 million times a year?
Something something something makes me think there are inefficiencies at play here.