Protecting Artificial Intelligence from Attacks using Quantum Computing

2023-08-15

Machine learning is a field of artificial intelligence where computer models become experts in various tasks by consuming large amounts of data. For example, modern AI for chess does not need to be taught chess strategies by human masters, but can independently "learn" these strategies through millions of games played against its own replicas.

In cases where it is not feasible or practical to write explicit instructions, this becomes invaluable - how do you define a mathematical function to tell you whether an image contains a cat or a dog?

Human children have never learned such a function, but instead develop an understanding of the differences between cats and dogs by seeing many examples of them.

Machine learning replicates this process in computers.

However, despite their great success and increasingly widespread applications, machine learning frameworks are still highly vulnerable to adversarial attacks - that is, malicious manipulation of their data that causes them to fail in unexpected ways.

For example, image classification models (which analyze photos to identify and categorize various objects) are often deceived by carefully crafted input images called perturbations, which are so small that they are imperceptible to the human eye.

Machine learning neural networks remain highly susceptible to such attacks, raising serious concerns about the security of deploying machine learning neural networks in situations where lives may be at stake. This includes applications like autonomous vehicles, where the system may be confused by graffiti on an otherwise harmless stop sign, causing the car to drive through an intersection.

At a critical moment in the rapid development and deployment of artificial intelligence, our research team is investigating how to use quantum computing to protect AI from these vulnerabilities.

Machine Learning and Quantum Computing

The latest advancements in quantum computing have generated great excitement about the prospects of enhancing machine learning with quantum computers. Various "quantum machine learning" algorithms have been proposed, including quantum generalizations of standard classical methods.

Generalization refers to the ability of a learning model to correctly adapt to new, previously unseen data.

It is believed that quantum machine learning models can learn certain types of data faster than any model designed for current or "classical" computers.

Conventional computers process data as either "0" or "1" - a two-level classical system.

Quantum computers use "qubits," which are the states of two-level quantum systems, and they possess peculiar additional properties that allow them to solve certain problems more efficiently than classical systems.

However, it is unclear how widespread these speedups will be and how useful quantum machine learning will be in practice.

This is because while quantum computers are expected to efficiently learn a broader range of models than classical computers, it is not guaranteed that these new models will be useful for the majority of machine learning tasks that people are genuinely interested in. These tasks may include medical classification problems or generative artificial intelligence systems.

These challenges have prompted our team to consider what other benefits quantum computing can bring to machine learning tasks, in addition to improving efficiency or accuracy.

Protecting AI from Attacks

The latest research published in Physical Review Research suggests that quantum machine learning models can better withstand adversarial attacks generated by classical computers.

Adversarial attacks work by identifying and exploiting the features used by machine learning models.

However, the features used by general quantum machine learning models are inaccessible to classical computers, making them invisible to adversaries who only have access to classical computing resources.

By simultaneously using classical and quantum networks, these ideas can also be used to detect the presence of adversarial attacks.

In normal circumstances, the two networks should make the same predictions, but in the presence of an attack, their outputs will diverge.

While this is promising, quantum machine learning still faces significant challenges. The most significant of these is the enormous gap in capabilities between traditional computing hardware and quantum computing hardware.

Today's quantum computers are still severely limited by their size and high error rates, preventing them from performing long computations.

Formidable engineering challenges still exist, but if these challenges can be overcome, the unique capabilities of large-scale quantum computers will undoubtedly bring transformation to a wide range of fields.

RECENT AI NEWS

RECENT AI TOOLS