Digital Security
The start-up Skyld joins the Berkeley accelerator
Date:
Changed on 10/06/2024
The Rennes-based start-up Skyld aims to improve the security of artificial intelligence algorithms embedded in telephones and other smart objects. After a year of development in the Inria Startup Studio, the project was recently selected by SkyDeck, the accelerator at the University of California, Berkeley, one of the most selective of its kind in the United States. As Skyld founder Marie Paindavoine explains, access to this internationally renowned programme will accelerate the market launch of protective technology that is urgently needed among AI publishers.
Because of their size, deep learning algorithms have long been the preserve of cloud computing. They ran behind the scenes on well-protected servers. Then, over time, these algorithms became smaller.
Founded by Marie Paindavoine, a cryptography scientist, the start-up designs innovative methods to mask the source code of deep learning models and hide information during execution by the processor. The project spent a year maturing with Inria Startup Studio, a programme, which helps researchers and engineers set up new companies based on research results.
That was when AI began to migrate to mobile phones and the Internet of Things, but this gave rise to a new danger: malicious people can physically access the device hosting the application and therefore try to reverse engineer it to extract the learning model and use the AI for their own purposes.
In other words: outright theft of intellectual property. For the victims, the loss is not limited to the amount initially invested in training the model. The companies affected also risk losing their technological lead to unscrupulous competitors, hence the need to protect AI algorithms in these insecure environments. This is precisely the aim of Skyld.
Automatically extracting AI models
Last June, Skyld gave a high-profile presentation at the latest edition SSTIC, a cybersecurity conference held every year in Rennes. Marie Paindavoine and two colleagues unveiled ModelHunter :a research prototype capable of analysing Android apps on PlayStore in order to automatically extract the deep learning models. The demonstration was an eye-opener for many participants.
Most companies are able to protect themselves against static attacks, but they’re not immune to extraction by dynamic attack. That’s exactly what we want to demonstrate.
The presentation was based on a research article published in the conference proceedings. Some of the work, however, was not disclosed because of the sensitivity of the subject and the risk of such a tool falling into the wrong hands.
Direction : Silicon Valley
Meanwhile, Skyld was also recently selected to join the European section of SkyDeck, the start-up accelerator at the University of California, Berkeley. “It is a highly selective programme. Out of 1,800 applications, only 33 were accepted”, says Marie Paindavoine. The entrepreneur has just returned from a whirlwind month in Silicon Valley. “I probably met nearly 150 advisers, mentors and investors during a “meet and greet” session with a 20-minute turnaround.” Incidentally, the entrepreneur even got a chance to give a legendary elevator pitch. “In theory, it’s a one-minute speech during which you present a project to someone in the time it takes to ride an elevator. Well, it actually happened to me, as I was ascending to the 13th floor of the SkyDeck building!”
The second part of the support programme is due to take place in Milan. It also includes funding of €145,000, in return for which SkyDeck will take an equity stake. “Technically, it’s a BSA-Air, or what is called a SAFE* in the USA.” This equity subscription is a fast investment agreement that allows the investor to convert their contribution into shares at the next capitalisation event.
In parallel, Skyld recently landed its first customer. HopeValley AI is another start-up developed by Inria that has designed an artificial intelligence solution to detect the early stages of breast cancer based on weak signals. Other contracts could follow shortly, including “a big name in software publishing”.