Can AI Defeat Image Obfuscation?

It's easy to obfuscate an image beyond recognition. But what if you want to discriminate only against robots?

Much like humans, AI doesn't tend to "un-grow". Academic studies are shared, experiments are conducted, and the capability of your average artificial neural network, swells, year on year. The field of automated image analysis has broad applications, such as reconstructing images after common obfuscation techniques, such as pixilation or blurring have been applied. The ability of artificial intelligence to make sense of images has improved a great deal in recent years, as can be demonstrated by uploading an image to Googles Vision AI.

I recently wondered how well a tool like Vision AI might cope with obfuscated images. Sure, if you don't want an image to be recognizable to AI, you can use any number of tools, to destroy any semblance of the original, but what if you want an image that makes sense to humans, while mystifying AI? One application for this need would be in a HOSP, or a Human-Only Solvable Puzzle, a bit like a CAPTCHA.

So let's have a look at what Vision AI can do, with a photo I took of the wonderful @eriluks...

How would you describe this photo? Your description most likely weighs the significant elements perfectly, and even infers some meaning. Let's see what Vision AI thinks.

Great job at analyzing the objects, but what about the story/meaning?

I don't know about you, but I'm impressed. I tried this with a number of images, and it did really well almost every time. As you can see below, it can also analyze facial expression to conclude that "confidence" is the subjects mood. On it's own or even as part of a sequence, this image would surely be very easily "understood" by AI.

If the application is a Turing Test or HOSP, it would surely be ideal to use images that were more easily understood by humans, while remaining mysterious to AI. These semi-abstract images might have to be generated on the fly, or encrypted at the very least, as any image online for longer than a few minutes will likely get the full Google spider treatment, including SEO analysis of related text, links, etc.

Below you can see some examples of distortions I made to the image. Keep in mind, the idea is to throw off images analysis while keeping the image coherent to humans. Techniques I tried were pixilation, line art, painting style, and Photoshops Cutout Filter. I've presented them in that order.

Although the cutout filter threw AI the most, it also pushes the limits of human perception. Still, having only tried a handful of options out of an endless number of obfuscation methods, I'm hopeful that automated image obfuscation could be used to effectively discriminate against AI, at least for now!

Further reading:
AI Can Recognize Your Face Even If You’re Pixelated
Vision AI
AI, Decentralization And The Captcha Arms Race
Proof Of Person Blockchain Project