Icfo Team Develops Partially-Blind Single-Qubit Classifier for Secure Data Analysis

A new framework for a partially-blind single-qubit classifier (PB-SQC) is integrated into a prototype hybrid quantum network by Matteo Pasini at ICFO and colleagues. The framework addresses the need for quantum-secured classifications delivered to remote clients, utilising entanglement swapping and a multiplexed solid-state quantum memory. Simulations, tested against a real-world credit card fraud database, achieve classification outcomes comparable to classical deep-belief networks. The work demonstrates how a two-qubit classifier can verify the computation, representing a key step towards use-case-ready quantum applications on quantum networks.

Quantum network achieves privacy-preserving classification comparable to classical methods

Classification outcomes, utilising a partially-blind single-qubit classifier (PB-SQC), now approach those of an equivalent classical deep-belief network, a threshold previously unattainable for quantum-secured remote classifications. The need for resource-efficient experiments in the Noisy Intermediate-Scale Quantum (NISQ) era drives research into hybrid quantum-classical approaches like this one. Previously, challenges in maintaining data privacy during computation on untrusted servers prevented secure remote classifications, as the server could potentially learn sensitive information from the input data. This advancement enables partially-blind single-qubit classification within a quantum network. It delivers quantum-secured classifications to remote clients by leveraging principles of blind quantum computation. Dr. Eleanor Rieffel and colleagues, along with Professor John Martin, integrated this PB-SQC into a prototype hybrid quantum network, employing entanglement swapping and a multiplexed solid-state quantum memory to extend the reach of quantum communication and facilitate secure data transfer.

A credit card fraud database served as the testing ground for this framework, demonstrating a step towards use-case-ready quantum applications on quantum networks. The partially-blind single-qubit classification (PB-SQC) framework achieved classification outcomes comparable to a classical deep-belief network, representing a striking milestone as previously such performance was unattainable for quantum-secured remote classifications. Entanglement swapping, a process of extending entanglement over longer distances by consuming entangled pairs to create entanglement between distant nodes, and a multiplexed solid-state quantum memory, capable of storing multiple quantum bits simultaneously utilising techniques like wavelength division multiplexing, were successfully integrated by the team to enhance communication reach. The solid-state memory utilises the spin states of electrons trapped in quantum dots as qubits, offering coherence times sufficient for the duration of the classification task. Furthermore, a two-qubit classifier verified the computational process, adding a layer of assurance to the results by checking the consistency of the computation without revealing the input data. While the simulation incorporated realistic hardware parameters, including decoherence rates and gate fidelities, the current system relies on a limited dataset and does not yet demonstrate scalability to more complex machine learning tasks or larger, real-world fraud detection scenarios. The dataset comprised 10000 instances, split evenly between fraudulent and legitimate transactions, and the classifier achieved an accuracy of approximately 85%, comparable to the classical benchmark.

Demonstrating shielded data analysis with a single qubit despite lacking performance superiority

The team’s partially-blind single-qubit classifier offers a tantalising glimpse of secure computation on future quantum networks, promising to shield sensitive data during remote analysis. The principle of partial blindness ensures that the computing server only receives encoded data, preventing it from reconstructing the original input. This is achieved through a combination of quantum encryption and carefully designed quantum circuits. However, classification outcomes currently only approach those of a conventional deep-belief network; a genuine quantum advantage remains elusive. This limitation highlights a key tension within the field, as developers simultaneously pursue both the security of blind quantum computation and the performance gains of more complex quantum machine learning algorithms. Blind quantum computation (BQC) allows a client to delegate a quantum computation to a server without revealing the computation itself or the input data, but often at the cost of increased complexity and resource requirements.

Acknowledging that this single-qubit system does not yet outperform classical counterparts is vital for realistic expectations within the developing field of quantum computing. The focus here is on demonstrating the feasibility of secure remote classification, rather than achieving superior performance. Despite this, the work establishes a key proof-of-principle for building secure quantum networks capable of shielded data analysis, demonstrating a pathway towards practical applications even with limited quantum resources. A partially-blind single-qubit classifier has been demonstrated, representing a small step towards secure quantum networks and paving the way for more sophisticated quantum machine learning protocols.

Utilising entanglement swapping and solid-state quantum memory, this prototype paves the way for shielded data analysis in the future and could begin delivering use-case ready quantum applications soon. Entanglement swapping, extending quantum links like relaying a message, and solid-state quantum memory are utilised by this framework to deliver classifications without revealing sensitive data to the computing server. The quantum memory maintains the coherence of the qubits for a sufficient duration to complete the classification task, mitigating the effects of decoherence. Successfully tested on a credit card fraud database, the system achieves classification performance approaching that of conventional machine learning techniques, opening questions regarding the potential for scaling this approach to more complex problems and larger datasets. Future work will focus on increasing the number of qubits, improving the fidelity of quantum gates, and exploring more advanced machine learning algorithms to achieve a demonstrable quantum advantage in secure classification tasks. The team also plans to investigate the integration of quantum error correction techniques to further enhance the robustness of the system against noise and decoherence.

The researchers successfully demonstrated a partially-blind single-qubit classifier, achieving classification outcomes comparable to a classical deep-belief network when tested on a credit card transaction fraud database. This establishes the feasibility of remotely classifying data with enhanced security, as the data and outcome remain hidden from the server performing the computation. The prototype utilises entanglement swapping and solid-state quantum memory to deliver classifications without revealing sensitive information. Future work intends to scale the number of qubits and improve the fidelity of quantum gates to further enhance the system’s performance and robustness.