One of the protagonists of Zero Historya 2010 novel by William Gibsonwears “a horrible t-shirt” with a strange design printed on it that prevents cameras equipped with artificial intelligence from recognizing him as a person: “The camera sees him, but then forgets that he saw him”, in the words of the writer, one of the fathers of modern science fiction.
A decade later, what was only a fictional hypothesis has become reality: for some time now, humans have been fighting a battle with machines for the protection of their privacy and identity, and today this type of clothing can already be bought and worn. Also in Italyproduced and made with Italian skills.
Interview
The future seen by Annalisa Barla, the teacher who teaches machine learning
by Emanuele Capone
Cap_able and adversarial patterns
“Biometric detection devices are everywhere, especially in countries like the United States, China, Japan (and soon here too, ed): they recognize faces, they acquire our data and our identities without us even realizing it and without knowing where this information goes “, he told us Rachele Didero, CEO and co-founder of Cap_ablea startup born in July 2022 that precisely produces and sells clothing that prevents these operations.
Didero has got the idea in 2019 during a period of study in New York, and the secret lies in what the Americans call the “adversarial pattern”, and which in Italian could be translated as an adversarial image: “Our solution allows you to hide without hiding, that is, without having to cover the face, which is illegal – he explained to us Federica Busani, the other co-founder – And indeed, with our clothes you stand out among the others and you are strongly recognizable from other humans, that is, from our species, but unrecognizable to machines ”. Who see us but forget they saw us, to put it like Gibson.
Busani e Didero they are both less than 30 years old, they are from Piedmont and Cap_able was born from their meeting: “We are an innovative benefit startup, we started with a contribution of 5 thousand euros from the Region and we financed by themselves with another 20-25 thousand euros“. Money that was mainly used for research and development, that is, to find the right images to print on t-shirts, sweaters, sweatshirts, pants and dresses.
How adversary clothing is created
The two founders have patented the fabric used for theirs Manifesto Collection in February 2021also thanks to a partnership with the Polytechnic of Milan (among the partners of the startup there is also Professor Giovanni Maria Conti), but getting there was not easy: months of studies on the machine learningsui fabricson the volumes of the body to understand which ones colors and shapes combine together to hinder the activity of AI engaged in facial recognition.
“We tested with object and person recognition software, such as Yolo (which is this, ed) – Didero told us again – to find out how to mislead them. Then we made the first samples of what would become our clothing, to be able to conduct even more in-depth tests “. To date, Cap_able dresses are made starting from 14 colors, two different types of shapes (geometric or rounded) and above all on specific fabrics: “The more reflective ones, such as cotton, viscose and silk, are more effective for the purpose we have given ourselves – said Didero, who also deals with the design of the garments – Better not to use wool, mohair or cashmere, if not in small percentages and in any case mixed with the rest “.
That the combination of some colors (very intense yellows, shades of red and green) and the arrangement of specific shapes puts the machines in difficulty is nothing new: in March 2019, the University of Ku Leuven in Belgium published a study explaining just this and demonstrated the effectiveness of what was discovered online. Busani, who is also a business developer of Cap_able, confirmed that “the drawing is somehow stronger than the person’s face, it stands out and confuses the recognition software” and that “the machine just doesn’t understand that behind that drawing there is ‘it’s a person”. What do you mean? “When an AI understands that what it is looking at is a personbegins a more in-depth analysis and searches for details such as facial features, eye position, facial expression and so on: our clothes block this process even before it begins “.
youtube: the Ku Leuven University experiment
How to buy Cap_able clothing
The startup’s garments are made in Carpi, in the province of Modena, in the Italian headquarters of the Japanese Shima Seiki, one of the leading companies in the knitwear sector: at the moment you can see and book online (leaving a deposit of 49 euros)but will soon be available on Kickstarter as well.
“On September 27th a crowdfunding will be active that will go on for a month and that we will use both to raise funds and to start production”, explained the two founders of Cap_able: the page is already accessible but not yet active, and the initial goal is to raise 80 thousand euros. The offer should be divided into three: 15% discount on the final price for supporters, 25% for the so-called early birds (i.e. the first to join) and 40% for VIPs, who should be those who already have left a deposit on the site. In the intentions, the shipment of the garments should be completed starting from December 2022: they will be 250 divided on 2 models, while the entire collection (consisting of 7 models) it will be completed in the course of 2023. Furthermore, the idea is to also create a line of accessories, such as bags or hats, to verify the effectiveness of these designs even with smaller surfaces available.
I prices are not popular: price list, 380 euros for a t-shirt and 560 for a hooded sweatshirt. There are some reasons: “We must take into account the time invested in research, which has a cost – we were reminded by Cap_able – and the fact that all production is made in Italy and using quality raw materials, such as Egyptian long staple cotton. that complies with the guidelines of the Better Cotton Initiative“. Not only that, we add: the value that is given to one’s privacy and the protection of one’s image must be considered. Specifically, of one’s face.
The business of the market of faces
One thing to understand, when it comes to identification and facial recognition, is the use that is made of people’s faces, because the problem is not only what happens in the moment, but also what will happen next: there are companies, such as the American Affectiva or Clearview AI, who base their business on databases of millions and millions of faces. They use them (among other things) to train artificial intelligences to understand the expressions of humans and they collect them just like that, by scraping the images captured by the camerasphotos posted on social networks or online in general, including on information sites, YouTube or other video upload platforms.
Not only that: there is also who uses these faces to create deepfakes (What are?) best and practically indistinguishable from the originals, to make red light movies, to impersonate a famous person or for other activities bordering on legality or even completely illegal. And it definitely is possible that even our face, perhaps captured by the apparently harmless camera of an airport or a shopping center, is used for these purposes. Which is why we should be careful that it doesn’t end up in the wrong hands or otherwise beyond our control, as we mentioned a few weeks ago on Italian Tech.
The case
People believe deepfake faces more real than real ones
by Carlo Lavalle
The ethical question and the future
Legality is also an important point for the two young founders of Cap_able, and when we asked them if they are not afraid of a criminal use of their clothing (for example, to rob a bank and not be recognized by the cameras surveillance) gave us an interesting answer: “Our clothes are not an invisibility cloak, the images are not altered or blurred. The people filmed can be seen, and you can see that they are people. It is only the AIs that do not see them, but if it were necessary to identify or recognize someone, it could still be done. But another person should do it ”. The point is precisely this: “What we want is to prevent is the indiscriminate collection of faces, faces, data and identities by these biometric recognition systems, which are too many and too invasive. And also to push a little to reflect on the space we are giving to these technologies ”.
Technologies that will become better with the passage of time, because the same adversary images used to deceive them are also used to train them to overcome these obstacles: “But we too will become better – Busani and Didero told us with a smile – Already today we are studying new combinations of colors and new materials to be inserted into fabrics, such as those that change according to light or temperature, which can prevent AI not only to recognize the face, but also the movement“. In a constant race to prevail, which perhaps not even William Gibson would have imagined.