New AI Can Tell Authentic Paintings And Fake Ones Apart

A Massachusetts couple, Steven and Andrea Frank, have programmed a new AI algorithm that can detect art forgeries. They described that their study, which is under evaluation at IEEE Transactions on Neural Networks and Learning Systems, is based on the entropy concept in thermodynamics and information theory.

Steven Frank, who is a member of IEEE, full-time patent attorney and part-time AI coder, said about this:

Entropy won’t be of much use for critics. However, it could be for a computer. Frank argues that image entropy may be the solution to a longstanding problem in processing high-resolution images using AI.

What the couple did was training a “convolutional neural network” or an AI often deployed in image recognition algorithms. The training included hundreds of works of Dutch artist Rembrandt and known fakes.

Here was when entropy came in handy. Instead of feeding the AI with entire paintings’ complete scans, that are massive files in gigabyte-size ranges. It could take forever for these files to be processed. With the couple’s method, the scans are split into smaller tiles, with a total of 13,000 ones.

One surprising result that the couple got from the analysis was working out which image parts were the most important in identifying whether a painting is real or not.

In order for the AI to be effective, a zoomed-out view is still required.

Frank said:

The initial test yields promising results as the neural network succeeded in differentiating real paintings from fake ones with a success rate of 90.4 percent.