What if, every time you were about to post a photo, a virtual assistant could scan it and tell you if anything in the image could compromise your privacy? I wonder how it would change the way we think about what we share on social media.
This paranoid BonziBuddy could send up a red flag in obvious cases, like when it sees you're about to post an image of your credit card or driver's license. But it could also remind you that you might be letting people know where you work if you're about to post a desk selfie. Or, it could warn you if a photo contains a clear shot of your fingerprints or an identifying tattoo.
Far from being annoying—OK, it might be a little annoying—this sort of tool might actually remind us of the many gray shades of personal privacy that we navigate every day on the internet. It might have helped the Russian soldier in the Ukraine who was tracked by his selfies, or the two Quebec women who plead guilty to smuggling cocaine into Australia and documented their whole trip on Instagram.
Although speculative, this sort of tool isn't strictly science fiction. A paper from researchers at the Max Planck Institute for Informatics in Germany, posted over the weekend on the arXiv preprint server, describes a new kind of "visual privacy advisor" that can look at photos and give them a red, yellow, or green light, privacy-wise. This is helpful, because while many people adjust their privacy settings to control who can see what they post, the posted content itself might be an afterthought.
The key to this system, the authors write, is deep learning. Deep learning "trains" a computer program known as a neural network to look for certain patterns in a large number of images. That way, when it looks at a new image, it can pick out the same patterns.
The researchers first had to create a training database. To do this, they collected tens of thousands of images and annotated them with 68 privacy categories: religion, occupation, identifying tattoos, addresses, hair colour, and so on. Then, to understand which categories are most important, they ran two surveys on 305 people. First, they asked respondents to rate how violated they would feel if they accidentally shared an image related to each of the 68 privacy categories. The participants were then shown photos (but not told which of the privacy categories they related to), and asked how comfortable they would feel sharing the image online.
Basically, the surveys asked how people ideally feel about their privacy boundaries versus how they enact them in practice.
After training the neural network on these sources, the researchers had a program that could look at an image and make a prediction as to the potential for privacy violation. Interestingly, the authors found that the survey respondents didn't always practice what they preached, which is exactly why their little tinfoil hat-wearing Clippy is necessary, they argue.
"The significance of this research direction is highlighted by our user study which shows users often fail to enforce their own privacy preferences when judging image content," the authors write. (They could not immediately be reached for comment.)
The system has a long way to go, however, the authors note. For example, while the network excelled at identifying privacy violations related to faces, hair, etc., as well as scenes (airports and hospitals), it couldn't reliably tell the difference between, say, a student ID and a driver's license. As with most things related to AI, this could be overcome in the future with more specific training on these problem objects.
While clearly an early step, it might not be long before Siri asks, "Are you sure you want to post that vacation selfie?"
Subscribe to pluspluspodcas t, Motherboard's new show about the people and machines that are building our future.