Secure Quantum Encryption Protects Data during Remote Neural Network Training and Use

Researchers are addressing the critical challenge of data security in cloud computing, where protecting sensitive information during remote computation is paramount. Sergio A. Ortega from the Departamento de F ısica Te orica, Universidad Complutense de Madrid, and Miguel A. Martin-Delgado, affiliated with both the Departamento de F ısica Te orica, Universidad Complutense de Madrid, and the CCS-Center for Computational Simulation, Universidad Polit ecnica de Madrid, demonstrate the first practical implementations of a perfectly-secure homomorphic encryption (QHE) scheme applied to neural networks. Their work utilises efficient Clifford+ decomposition to implement convolutional neural networks for two key scenarios: reverse delegated training, enabling federated aggregation of encrypted data from multiple providers, and private inference, allowing users to process encrypted data with remote networks. Significantly, this analysis of server circuit privacy reveals probabilistic model protection via Pauli gate concealment, establishing perfectly-secure QHE as a viable framework for multi-party computation.

This breakthrough addresses a critical need for data privacy in cloud-based quantum computing, where sensitive information must be protected during remote computation. reverse delegated training and private inference.

Reverse delegated training involves a novel approach where a user possessing a quantum neural network lacks the necessary training data and instead receives encrypted data from a provider, reversing the typical delegated training paradigm. Simultaneously, the study showcases private inference, enabling users to process encrypted data with remote quantum networks without revealing their information to the network owner.

These applications diverge from conventional delegated computing models, where resource limitations typically drive the need for remote execution, instead highlighting scenarios where data or network ownership is distributed. These implementations move beyond theoretical concepts, utilising realistic simulations where sensitive data resides solely in user-provided quantum states, mirroring practical applications.

Furthermore, analysis of server circuit privacy reveals a probabilistic model protection mechanism through Pauli gate concealment, addressing the crucial concern of safeguarding the network’s internal structure during decryption. The efficient implementation, calculated in terms of T-gate complexity, demonstrates the potential for scaling these techniques to more complex algorithms and datasets.

Researchers implemented quantum convolutional neural networks (QCNNs) to demonstrate the feasibility of these applications, explicitly calculating computational complexity in terms of T gates, a key metric for assessing quantum algorithm efficiency. Reverse delegated training scenarios demonstrated a successful implementation where encrypted data from multiple providers trained a user’s network via federated aggregation, achieving fault-tolerant QCNN training with a total gate count of 1,123,789 T-gates for a network operating on six qubits.

This complexity confirms the feasibility of implementing such networks within the constraints of the chosen homomorphic encryption scheme. This indicates a strong level of protection against adversaries attempting to extract information about the network’s internal parameters during the decryption process. The research further quantified the leakage rate of information about the network’s weights, finding it to be less than 0.05 bits per gate operation.

This low leakage rate is crucial for maintaining the confidentiality of the network model. Realistic simulations, performed on algorithms embedding sensitive information solely in user-provided quantum states, showcased the practical applicability of the perfectly-secure QHE scheme. This work prioritised realistic implementations, moving beyond theoretical demonstrations to address practical challenges in cloud-based quantum computation. The chosen processor facilitated the construction of QCNNs for reverse delegated training and private inference, both designed to protect data confidentiality.

To enable reverse delegated training, where encrypted data is supplied by one party to train a network owned by another, circuits were designed to accommodate federated aggregation of encrypted information. This approach focuses on a collaborative scenario where data ownership is separate, diverging from conventional delegated training. The implementation leveraged efficient Clifford+ decomposition to optimise circuit depth and fidelity.

Private inference was realised by allowing users to process encrypted data with remote networks, maintaining data privacy even when the network is held by a third party. Scientists have achieved a significant milestone in practical quantum cryptography, demonstrating realistic implementations of perfectly-secure QHE applied to neural networks. For years, the promise of processing encrypted data, homomorphic encryption, has been hampered by computational overheads.

This work bypasses those limitations through efficient quantum circuit design, specifically leveraging Clifford+ decomposition to make calculations feasible on near-term quantum hardware. The focus on convolutional neural networks is astute, as these are ubiquitous in modern machine learning applications, from image recognition to natural language processing.

The ability to train models on encrypted data from multiple sources without revealing individual datasets addresses growing concerns about data privacy and collaborative machine learning. Similarly, secure inference allows users to query remote networks without exposing their sensitive inputs. The probabilistic model protection afforded by Pauli gate concealment adds another layer of security, though its effectiveness requires further scrutiny.

However, the current implementation remains constrained by the limitations of available quantum hardware. While the circuits are efficient by quantum standards, they still demand a substantial number of qubits and gate operations. Scaling these systems to handle more complex neural networks and larger datasets presents a formidable challenge.