Artificial Intelligence and the Detection of Forgery
Hephaestus is the world’s only company that can deploy artificial intelligence to identify artworks with greater than 98.2% accuracy. The results are so conclusive that we are working with the world’s largest insurance companies to issue the first fraud insurance product on paintings. While the implications of this development are profound, in this article we would like to focus briefly on Hephaestus’ use of artificial intelligence algorithms to solve classic problems in art history and attribution -- one significant way in which Hephaestus distinguishes itself from all its competitors.
Art forgers are often underestimated, and sophisticated forgers can be amongst the most knowledgeable people on the planet with respect to technology and technical art history. Scientific tests can be reverse engineered: historic pigments can be recreated or reused, and carbon dating measurements can be contaminated and manipulated. Yet, these methods continue to be used by the art market, without any regard to addressing key issues.
In addition to the misuse of scientific tests, the art world’s reliance on connoisseurship and expertise is also problematic: let’s face it, the “referees” in the art market – the experts who declare a painting authentic have powerful incentives and conflicts of interest. As a result, the art markets are riddled with scandal after scandal, but the duopoly of the major auction houses exerts a kind fascination on those who willingly suspend their disbelief. In fact, Sotheby’s and Christies once put the very same painting by the very same artist up for sale at the very same time: a still life attributed to Gauguin.
The situation is analogous to what occurred in the Global Financial Crisis of 2009. It was not so much a liquidity crisis as it was a crisis of trust. The “referees” in this case were the rating agencies, Standard & Poor’s and Moody’s, who were paid by the investment banks to rate collateralised loans. When the referee said “yes” everyone at the table was off the hook: investors got to buy AAA rated securities, investment bankers made their fees, rating agencies were paid, and unqualified buyers got to borrow money.
We categorically predict that the compromised position of authenticators of the art market will cause the same meltdown as the digital age increases transparency and “Made You Look” documentaries expose the opacity and pathological “secrecy” of the market.
Hephaestus was founded with the mission of eliminating forgery, which is only possible with the removal of shortcomings of scientific tests and conflicts of interest. By applying stringent forms of testing for the physical, material aspects of paintings, and applying them in a scientific sequence such that each test increasingly eliminated, a far more accurate and reliable set of scientific protocols was developed. Nonetheless, scientific study of materials can only ever prove something as a forgery. The determination of authorship and attribution has always been a thorny subject.
Enter machine learning and Hephaestus’ crossover with astrophysics. The mathematical inferences of large data sets of light have given scientists a greater understanding about the origins of the Milky Way derived from the mathematical inference of large data sets of light. If astronomers could use artificial intelligence to develop “light signatures” from data sets about objects that were light years away, why couldn’t similar forms of analysis be used to develop “signatures” for the brushstrokes evident in paintings?
Working with preeminent astronomers at Princeton University, the University of London and the Australian National University, Hephaestus used “scattering transform” to detect more than one million dimensions and reduce them to about a 100-dimensional “fingerprint”. Some of these dimensions include the pressure with which a paintbrush is applied to a canvas, the average distance between objects, the colours themselves, the spacing of strokes, and the presence or absence of shapes in the artist’s oeuvre. If you will, Hephaestus' Featuriser extracts these characteristics for each artist, assigning each one a value that assess its contribution to the artist’s style. A three-dimensional contour results and the degree to which a painting “fits” inside the contour determines the likelihood that it may be attributed to an artist. This methodology allows for explainable Artificial Intelligence. It is not a black box. All its results may be traced and explained. And these are the stringent protocols and methods that set Hephaestus apart from its competitors.
It is not only our competition that cannot imitate us; neither can forgers. Two critical outcomes of Hephaestus’ tech are: (1) these dimensions are not visible to the eye and may only be detected by statistical analysis of millions of data points concurrently (2) they are therefore nearly impossible to forge. We have not yet dealt with the case of a living artist “forging” their own works.
An illustration may demonstrate the effectiveness of the technique: A classic problem in art history and connoisseurship is distinguishing the works of Canaletto from his nephew and apprentice, Bellotto. Hephaestus’ algorithms solved the problem with 99.2% accuracy.
The “devil in the details” is the collection of accurate data. For the use of forged works or misattributed ones will dramatically lower the accuracy of what may be detected. It is here that Hephaestus distinguishes itself from all its competitors because its databases use specialised photographic equipment on only authenticated works. The database of artists is growing slowly and painstakingly, but the results are astonishing. These are the highest evidentiary standards in the world.
So the next time a laboratory assures you “we use AI techniques” ask if they deploy Explainable AI or if it’s a black box solution and stand back.
Hephaestus: eliminating forgery from the art markets, one painting at a time.