Owl pics? Here’s how Tumblr censor bots are being fooled
Update (December 09):
For all latest updates on the ongoing Tumblr adult content ban episode, head here.
Update (December 09):
We have stumbled upon two more platforms – Explicitr and Suffra – that are also pitching themselves as Tumblr alternatives. More info on them here.
Original story follows:
From its iOS app getting deleted to the company announcing a sort of blanket ban on porn, a lot has happened over at Tumblr in the past few weeks. Now that the dust has settled down, we’re seeing users have either migrated to other platforms, or are looking for alternatives. Some are even launching petitions or planning protests.
However, it has now come to light that there are users who are trying to test the limits of Tumblr’s content censoring algorithm. Some screenshots and details posted in a blog post on Tumblr (which has now been deleted, but not before social media picked it up) reveal that the censorbots can be fooled by superimposing pics of certain size on the original content.
The following tweet shows how someone apparently managed to fool Tumblr’s algorithm by transplanting an owls image on a pic that was marked as explicit by Tumblr’s censoring bots. It’s worth mentioning that this was only achieved when the owl pic was scaled 50%. Take a look:
So here's something. Users are testing the limits of Tumblr's new algorithm that flags adult content (aka "censorbot"). This one found that a man's chest was flagged, but a man's chest next a 50% scale owl went unnoticed.
— Chappell Ellison ٩( ᐛ )و✎ (@ChappellTracker) December 7, 2018
There are many ways in which software’s detection of objects can be fooled and object transplantation is apparently one way. Someone on this very Twitter thread pointed towards a recent research on this:
An academic paper from August describes a technique called object transplanting. You can fool the algorithm by pasting other objects onto the image. https://t.co/yA7ys3YMIf
— … (@ultraspank) December 7, 2018
So what does this mean? Can owl pics save your NSFW content? Well, we aren’t sure at the moment, but even if that’s the case, Tumblr may tweak their algorithm to plug in this loop hole. For what it’s worth, chatter on Twitter suggests the company is using open_nsfw software from its parent Yahoo.
They're using open_nsfw, a neural network trained to identify nsfww content, which has a layer of two competing nsfw and sfw liking neurons that combine for the final result, so literally including a picture of a landscape with the image can hype up SFW Guy and lower the nsfw %
— Mincer the Chunk-Seeking Construct (@an_gremlin) December 7, 2018
It has been speculated that open_nsfw, the open source neural network developed by Yahoo (Tumblr's former parent company) is what Tumblr is using to flag content. An interesting article about such things:https://t.co/yA1FvN5dOP
CW for abstract NSFW art and, uh, math, I guess
— Michael Kinyon (PoGo: 7359 4209 7503) (@mkinyon) December 4, 2018
And as others have pointed out, the software has some major discrepancies.
According to Yahoo open nsfw, this is the most not safe for work picture that the mars curiosity rover has taken so far, with a score of 96%. pic.twitter.com/IEsimSEAlE
— Pim de Groot (@mifune) December 3, 2018
PiunikaWeb is a unique initiative that mainly focuses on investigative journalism. This means we do a lot of hard work to come up with news stories that are either ‘exclusive,’ ‘breaking,’ or ‘curated’ in nature. Perhaps that’s the reason our work has been picked by the likes of Forbes, Foxnews, Gizmodo, TechCrunch, Engadget, The Verge, Macrumors, and more. Do take a tour of our website to get a feel of our work. And if you like what we do, stay connected with us on Twitter (@PiunikaWeb) and other social media channels to receive timely updates on stories we publish.