Shutterstock’s new watermarking system foils Google’s AI
Following a study Google released last week showing how easily an AI could remove watermarks from stock images, stock photography website Shutterstock has introduced stronger watermarks that are more difficult to eliminate using automated software. Google notified the stock photography company about the vulnerability before publishing its research paper, and it seems the company took their warning in stride.
Google’s technology works via a computer algorithm that learns to identify the common elements in many stock images—that is, the watermark—and then remove just those elements without altering the underlying photo at all.
Google ultimately identified the consistency of watermarks as their main point of vulnerability. To counteract the vulnerability, watermarks needed inconsistencies that involved ‘random geometric perturbations.’ In other words: the marks should be warped to avoid consistency across the entire photo library. Doing this confuses the artificial intelligence, and results in watermark artifacts being left on the image after attempted removal.
Stock photography website Shutterstock has already implemented an improved watermark system based on Google’s advice. With its new system, stock photos are protected by watermarks that are applied with slight geometry changes, making it impossible for an AI like Google’s to figure out precisely which pixels belong to the watermark versus the image itself.
“The result was a watermark randomizer that our engineering team developed so that no two watermarks are the same,” says Shutterstock CTO Martin Brodbeck. “By creating a completely different watermark for each image, it makes it hard to truly identify the shape.”