Cybersecurity: malicious connected objects betrayed by their radio frequencies

Two smartphones or two smart watches of the same model and same brand are identical, aren’t they? “Don’t be fooled”, says Valeria Loscri, Head of Research of the Fun project team at the Inria Centre at the University of Lille. The standardised mass production of connected objects does not mean that there aren’t tiny differences or imperfections between them. These differences can impact the radio frequencies emitted by the objects, which can result in a shift in the carrier frequency or a phase shift on an oscillator or another part of the microcircuit. ”

The Fun (Future Ubiquitous Networks) project team therefore looked into a way of using these imperfections to effectively authenticate devices for security reasons. “They can be compared to identical twins. They look exactly alike but they don’t have the same fingerprints”, explains Valeria Loscri.

Unique and almost unforgeable imperfections

Scientists intend to use these imperfections to their advantage. Faced with the growing threat of illegal devices infiltrating wireless networks, to steal identity for example, the possibility of distinguishing a malicious device from others is a way to secure communication.

Realistic test subjects thanks to reprogramming

Using these actual measurements, the team simulated three realistic communication environment scenarios, involving a number of different objects, some of them authorised, and some not. The team experimented with these scenarios using software-defined radios, which are configurable emitting and receiving devices. Using these devices, they managed to reproduce the emissions of several connected object test samples, each one defined by its own specific and unique imperfections.

Machine learning to authenticate connected objects

These samples were analysed by training different machine learning models to recognise combinations of imperfections and characterise objects accordingly: Random Forest, Support Vector Machines, K-Nearest Neighbors, XGBoost and logistic regression.

“Of these models, Random Forest has proven the most effective”, says Valeria Loscri. “This machine learning algorithm associates the results of several decision trees to then produce one single output. This has been used to authenticate connected objects with a combination of three imperfections, with a detection rate exceeding 96%, even in the most demanding environments! ” 

The three scientists also tested their solution in various cyberattack scenarios, such as using interference to assess the resistance to background noise, or multi-node attacks involving several malicious devices acting simultaneously.

A robust, energy-efficient solution

Ultimately, this approach combining the identification of connected objects based on materials and machine learning models could prove more effective than traditional cryptographic methods when it comes to cybersecurity. Not only does the Fun project-team’s solution use much less computing power and energy, it also avoids the use of data encryption keys that can be intercepted or corrupted. 

What are the researchers doing now?

And the Fun project-team can applaud themselves for the results they have already achieved: “Our solution has proven to be effective, robust and flexible. A number of cybersecurity experts actually showed an interest in it at the Annual Computer Security Applications Conference (ACSAC) in December 2024 in Hawaii.” (See below)