This Algorithm Wants to Find Who??s Naked on the Internet
Source: Brian Barrett

The Internet recently fell in love with picture-parsing websites that guess your age, and whether someone is your twin. Now it offers another spin on novelty image recognition: an algorithm that knows if someone’s naked.

Isitnude.com, launched this week by a company called Algorithmia, delivers on the promise of its straightforward URL. Feed it an image―one it suggests, or one from your own menagerie―and it will tell you if anyone in that image is wearing little to no clothing. It is prudery, distilled into code.

But why does this even need to exist? Unlike the carnival game quality of how-old.net and twinsornot.net, “is it nude” seems easy enough to answer with our own eyeballs. It’s not the site that matters, though; it’s the algorithm behind it, and the service it can provide. Or rather, that it’s already providing.

“A customer came to us trying to run a site that needs to be kid-friendly,” says Algorithmia CTO Kenny Daniel. His company is an algorithmic clearing house, taking computational solutions from academia and beyond and offering them to the world at large for a fee. In this case, the customer wanted the ability to screen images with some confidence that they wouldn’t be pornographic.

From there, Daniel says, it simply was a matter of combining different bits of research Algorithmia already has undertaken or acquired. “We’d done quite a bit of work on image recognition,” Daniel says. “One of our engineers took some algorithms off the shelf, things like image detection, skin color detection, and then put them together.”

The algorithm builds on work that attempts to identify nudity simply by finding skin patches, which seems good sufficient until you realize a tight shot of, say, a hand is nothing but skin. To help weed out false positives, Algorithmia added a few layers of intelligence. It used its previous work in hand gesture recognition to fine-tune its skin tone smarts, and brought in-off-the-shelf face and nose detection algorithms to give skin tone range guidance and better compensate for multiple people in one image.

Algorithmia

And thus, its algorithm for nakedness was born. As for the website? Think of it less as a parlor trick, and more as a marketing tool. If sex sells, maybe filtering it will, too.

“We thought there was potentially much wider applicability,” says Daniel. “Anybody who’s trying to run a community but wants to filter out objectionable content, or keep it kid-friendly, could benefit from this same algorithm.” Given the nature of the content, Algorithmia doesn’t store or view your images; the site simply calls the nudity detection algorithm and spits back a result.

In the process, contributions to isitnude.com help fine-tune its detection prowess. That progress will be piecemeal, though; the privacy standards that keep Algorithmia from peeping your pics also means potentially helpful results need to be reported through other channels. “People have been emailing us, tweeting us, saying that a picture didn’t quite work out,” Daniel says. “We’ve been using that to iterate.”

It’ll also take more than just fine-tuning for the algorithm to become ironclad, or even lightly bronzed. It can’t yet recognize individual body parts, for instance, meaning that an image of someone who’s fully clothed except for, well, the parts you most want clothed, likely would still get through. And because it’s based purely on flesh, otherwise innocent beach pics might find themselves flagged. An even bigger hiccup? Because the algorithm relies on skin tone recognition, it’s powerless against black and white images.

There’s also the larger issue of its limited scope, even when it works flawlessly. There’s an entire universe of potentially offending image-based content out there that has nothing to do with skin, explains Joshua Buxbaum, co-founder of WebPurify, a content moderation service. “Violence,” Buxbaum gives as an example, “is complicated. We even get as detailed as violence in a sport that’s not a violent sport is violence, but if it’s boxing, which is a violent sport, we allow it.” That’s just for starters; a truly comprehensive image moderation suite, says Buxbaum, potentially needs to correct for everything from hate crimes to drug paraphernalia to rude gestures, depending on a client’s needs, and to address the gray areas within each of those categories. Algorithmia’s solution is only looking for naked people, and even then, still has the trying task of weighing whether they’re obscene or not. Buxbaum offers that the simplest algorithm could filter something for curse words, but it would have a difficult time identifying hate speech or bullying―meaning a nudity-spotting algorithm could find flesh, but have trouble distinguishing if it’s revenge porn or artwork.

That doesn’t render its nudity-detection algorithm useless though; it’s just a blunt instrument rather than a laser-guided prude missile. Daniel also expects these sorts of issues to be resolved over time, especially if there’s a clear commercial interest. “We’ll continue improving it,” he says, “or if a user in our system wants to come in and try to build a more sophisticated nudity-detection algorithm, they’d also be able to do that.” If one of Algorithmia’s partners did crack a better code, it would get paid in a model borrowed directly from Apple’s App Store; receive a royalty on every single algorithm call, with Algorithmia taking 30 percent off the top.

With a good enough solution, that could add up. Clearing out offending images from message boards or comment sections can be a full-time job, especially when trolls enter the picture. Being able to offload even some of that work to an algorithm would be a tremendous time-saver for even small sites; for a mega-corp like Facebook, which employs a small army of content moderators, it could be a real boon. Buxbaum says his company might even be interested at some point, but cautions that the system would need to be practically infallible: “If this nude algorithm was actually 99.9999 percent successful―and even that’s a little scary, because with nudity that .0001 percent is pretty petrifying, depending on who your audience―it would be something we were definitely interested in using.”

It’s not there yet, and won’t be any time soon. Still, Algorithmia’s ultimate goal is a system so advanced that it wouldn’t need a human backstop. “I’m definitely a strong believer that we will be in that world,” says Daniel, while also acknowledging that simply knowing if something is nude or not is only part of the equation. The much harder, if not impossible, part is determining what’s offensive to whom. “There are obviously going to be some gray areas that are difficult for any algorithm―or any human―to figure out on their own.”

Even a short amount of time playing with isitnude.net shows that we’re still a long ways off from confronting those philosophical content problems head-on. It can tell you how much skin there is, but without context (or body parts) that information doesn’t get you very far. But as the algorithm gets more sophisticated, the seemingly simple question of “nude or not?” will likely require even more human consideration, depending on who wants to know, and why, and what cultural mores they’re operating under. As Algorithmia now knows better than anyone, the things that look black and white are often the hardest to parse of all.