Prior to class this week I was listening to a presentation about how art collectors are now using AI to spot forgeries of artwork. Obviously, some pieces of art are very valuable because of the uniqueness and rarity of them. The ability to make obscene amounts of money from an “original” work of art breeds an abundance of copycats.
So how does AI actually do this? They look at paintings that have been validated to be done by a specific artist, like Picasso, and study all of the work that he has done over the years using high resolution pictures. All of the paintings have been traced back to the original artist through a verification process that analyzes the painting. The analysis of previous work is the most critical aspect of the whole process. Since any algorithm is “garbage in, garbage out” all of the inputs need to be perfect otherwise contamination could lead to a skewed output and actually cause more harm than good. The more comprehensive an artist’s career portfolio is, the stronger the neural network will be to determine the different styles of each individual artist down to their brush strokes. Using these other paintings to understand the style of brushstrokes a specific artist typically has, it develops a heat map for the painting to point out “areas of question” that indicate it might be a forgery.
This concept is not new and gained a lot of popularity 4 years ago when a paper was published by Rutgers and the Atelier for Restoration and Research of Paintings in The Hague that broke down 297 pieces of artwork by 4 major artists (Picasso, Matisse, Modigliani, and Hasse) into 80,000 strokes. The system detected the precise artist 80% of the time and fakes 100% of the time.
When a new, previously unseen artwork is being analyzed, the features are analyzed in comparison to the already stored one. If they match, the new image is labeled as original; otherwise, it is considered a fake. Probabilities to distinguish original from fake can be higher than 90% depending on the style and artist.
With any sort of machine learning, bias is a huge problem, which is why the algorithm is trained to learn the relevant features of an artist itself. The only feature given as an “input” is the brushstroke. It is not easy to find out what features have actually been learned, so the heat map is produced to provide a visual interpretation of the decision process.
This makes catching forgers a lot simpler for a variety of reasons:
- It runs on images alone so the work does not need to be transported. This is helpful because the original assumption is that this is a valuable piece of artwork, so transporting it physically opens up the opportunity for physical damage, loss, or theft.
- An inquiry can be responded to quickly as the algorithm can run in a few hours. Traditional committees that judge whether a painting is a forgery or not can sometimes take months to make a determination, with no more accuracy than the computer.
- The process is not invasive. Oftentimes, a sample is removed to do a chemical analysis to make sure materials used are similar to what was around in that time period, or compared to what an artist would normally use, but now that does not need to occur.
- Finally, there is more objectivity present in the use of a machine versus an individual’s opinion. Two experts could arrive at different conclusions based on less definitive or scientific information.
While spotting forgery is important and valuable work, this technology is also helpful in assigning who should be credited with a piece of work. Throughout history, artists might have the same theme multiple times or their students might have filled in paintings that were started and still given credit to the original for the idea.
Like everything else, this technology is a tool building confidence in identifying original artwork. As noted before, there are some downsides to AI and the algorithms can have bias if they tend to get too much information from a specific period in an artist’s life, particularly if their styles changed over time.
While some are skeptical of how far the technology can go, people will still need to have the final call in determining if a piece of art is authentic or not. However, we can be more confident in whatever verdict is determined. Additionally, this was tested on only a few artists within a specific time period. For other artists who have many strokes in a painting, using this technology will be more difficult as it means there is more information the computer will have to sift through to determine authenticity. Finally, there have been challenges with older paintings as well, as these might have been restored or overpainted several times.
I guess at this point, until blockchain becomes so ubiquitous that we can determine the history of a work from its inception, the only thing we need to worry about is the machine being smart enough to be able to create a perfect forgery itself…
For further Reading: