Research tests the potential of two quantum machine learning algorithms for malware classification

Over the past decades, cyber attackers have become increasingly skilled at compromising systems and circumventing security measures. As a result, detecting and pinpointing malware has become a pressing challenge for many businesses and individuals around the world.

Cybersecurity experts have recently been researching the potential of machine learning techniques to classify malware and determine what steps should be taken to eradicate it. While some of these techniques have achieved promising results, research has shown that many of them can be fooled or fail to accurately identify never-before-seen malware.

Researchers at Orange Innovation Inc. recently conducted a study evaluating the potential of quantum versions of machine learning algorithms in hopes of identifying more reliable methods for classifying malware. . their paper is arXivprovides initial insight into the strengths and limitations of two types of quantum machine learning models, and outlines possible directions to be explored in future cybersecurity research.

“I have been working on using artificial intelligence for malware analysis since 2019,” Tony Quertier, co-author of the paper, told Tech Xplore. “Together with Grégoire Valluet, who started my postdoc in October, I would like to explore what quantum technologies can bring to this problem. I have a background in science, so I hope to be able to use my theoretical knowledge to understand this subject. ”

Quertier and Barrué believe that quantum machine learning has the potential to help users extract more information from less data. To test this hypothesis in the context of malware classification, they so far evaluated the performance of two different quantum machine learning models known as QSVM and QNN.

“The first algorithm we tested was a simple QSVM adaptation of the support vector machine algorithm in Quantum,” explained Quertier. “Then we also tested QNN, a quantum adaptation of classical neural networks. The results have turned out to be very promising.”

Quertier and Barrué found in initial evaluations they conducted that the QSVM algorithm achieved very promising results, outperforming some of the team’s traditional SVMs for malware classification on several parameters. Did. QNN, on the other hand, optimized only by reuploading data and using a technique known as SPSB, was able to classify malware with 87% accuracy. Considering it was trained on a limited amount of data, this is quite good.

“Obviously, this accuracy is not as good as the previous version of the algorithm, but here we used only 1,000 samples, whereas the previous version was trained on 1 million data,” Quertier said. “For a first approach, this has exceeded our expectations. What is most interesting to me is the ability of quantum machine learning techniques to learn from limited training data. However, it is not so easy for some domains to hold large amounts of data.”

A major goal of Quertier and Barrué’s ongoing research efforts is to optimize algorithms to efficiently extract more information from limited amounts of data. In our next work, we plan to explore the possibilities of other quantum versions of machine learning algorithms, such as quantum convolutional networks (QCCN), while also using mathematics to optimize and better analyze available data.

“For example, using Lie theory, we can identify the number of parameters to achieve overparameterization (there are enough parameters in the model that the Fisher information matrix reaches maximum rank and yields maximum capacity). ), and even identify symmetries in the data, and adapt the quantum gates we use,” added Quertier. “In October 2023, a doctoral dissertation on this subject will be launched and will be supervised by Daniel Juteau, an expert in the field.” [type of] Math. ”