Quantum machine learning

Introduction

Quantum computers, when available, are expected to have a great impact on various computational problems. In particular, quantum machine learning is a very active research field in both academia and industry. A large number of quantum machine learning research groups were created in large companies, as well as a number of startups. Their goal is to develop a portfolio of algorithms in various areas such as finance or healthcare to be ready to use when quantum computers will be available.

Similar to their classical counterparts, quantum machine learning algorithms need to be trained on data. This paradigm raises two main security issues that already exist for classical machine learning. On the one hand, the algorithms themselves are produced by investments in research and development. Companies working in the area might not want to make these algorithms public. On the other hand, data required to train those algorithms might be subject to regulation that prevents sharing them.

Data security in a quantum world

Although these two security issues are similar for classical and quantum machine learning, these two setups have different constraints and which will lead to different solutions. Classical machine learning algorithms are resource consuming, and their execution is usually performed on specialized hardware. This hardware can be bought by big enough companies, but in the quantum case, this seems largely unrealistic because of the expected price of quantum computers. For this reason, most companies building quantum computers are also developing a cloud access to their hardware. How can the secret of the algorithm be maintained in such a setting?

Solutions from distributed quantum computing

Fortunately, there exists a solution coined as blind quantum computing, which can be achieved by connecting the quantum computer to a quantum communication network. In this setup, it is well-known that quantum programs can be efficiently and securely delegated to the server. The server then runs an encrypted version of the program and returns the answer without learning the computation it is running.

In the classical setting, this task is known as homomorphic encryption. Surprisingly, the quantum version is much more efficient than the classical one, and only induces a minor overhead on the computation as long as a quantum communication channel allows the client to drive the computation remotely. Moreover, it only requires a very light client which can be built with currently existing technology, and a conversion from the qubits used for transport to the qubits used for computing.

While blind quantum computing is often recognized as a killer-feature for quantum networks, we want to stress that it also fits well with the planned development of quantum computers and quantum networks. Quantum computers will be expensive and probably mutualized through HPC facilities accessible to multiple users. Creating quantum communication networks between those users and the facility will bring security for the algorithms.

Using quantum networks for data protection

The second issue that needs to be considered is the security of data. While startups may be developing quantum machine learning algorithms today, and get cloud access to quantum computers, getting access to sensitive data may not be possible today or in the future.

A solution to this problem for classical machine learning is to add noise to the data to preserve privacy. Machine learning algorithms are naturally resistant to noise (an algorithm that recognizes a cat on a picture should be able to recognize a cat on a noisy picture). Quantum networks being inherently noisy and lossy, we might be able to take advantage of that feature to enhance security and privacy.