Abstract:Secure multiparty computation enables the joint evaluation of multivariate functions across distributed users while ensuring the privacy of their local inputs. This field has become increasingly urgent due to the exploding demand for computationally intensive deep learning inference. These computations are typically offloaded to cloud computing servers, leading to vulnerabilities that can compromise the security of the clients' data. To solve this problem, we introduce a linear algebra engine that leverages the quantum nature of light for information-theoretically secure multiparty computation using only conventional telecommunication components. We apply this linear algebra engine to deep learning and derive rigorous upper bounds on the information leakage of both the deep neural network weights and the client's data via the Holevo and the Cramér-Rao bounds, respectively. Applied to the MNIST classification task, we obtain test accuracies exceeding $96\%$ while leaking less than $0.1$ bits per weight symbol and $0.01$ bits per data symbol. This weight leakage is an order of magnitude below the minimum bit precision required for accurate deep learning using state-of-the-art quantization techniques. Our work lays the foundation for practical quantum-secure computation and unlocks secure cloud deep learning as a field.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: during deep - learning inference, how to ensure the security of multi - party computation, especially when outsourcing computing tasks to cloud servers, how to protect the privacy of client - side data and model weights. Specifically, as deep - learning models' demand for computing resources continues to increase, many inference tasks are outsourced to the cloud for execution. However, this practice brings security risks and may lead to the leakage of client - side data and model weights. To solve this problem, the author proposes a quantum - secure multi - party computation method based on a linear - algebra engine. This method utilizes the quantum properties of light to minimize information leakage while ensuring computational accuracy. ### Main Contributions 1. **Information - Theoretic Security**: By using traditional telecommunication components, a linear - algebra engine is constructed to achieve information - theoretically secure multi - party computation. 2. **Deep - Learning Applications**: The linear - algebra engine is applied to deep - learning inference, especially the MNIST classification task, demonstrating its effectiveness in practical application scenarios. 3. **Information Leakage Analysis**: Through the Holevo theorem and the Cramér - Rao bound, the upper bounds of information leakage of model weights and client - side data are strictly deduced, and it is proved that these leakage amounts are very small, far below the precision required by existing quantization techniques. ### Experimental Results - In the MNIST classification task, the test accuracy rate exceeds 96%, while the leakage per weight symbol is less than 0.1 bit, and the leakage per data symbol is less than 0.01 bit. - The information leakage amount of this method is far below the minimum bit - precision required by existing quantization techniques, indicating that it has high security. ### Key Points of the Solution - **Photon - Computing Architecture**: Utilize the photon - computing architecture to achieve efficient optical matrix - vector multiplication. - **Quantum Properties**: Utilize the quantum properties of light to ensure the security of information transmission. - **Protocol Design**: Design a new protocol that enables the client to perform inference without exposing data and return a verification state to the server to check security. In conclusion, this paper solves the data and model privacy protection problems in the deep - learning inference process by introducing a quantum - secure multi - party computation method, providing a new solution for secure cloud deep - learning.

Quantum-secure multiparty deep learning

Quantum delegated and federated learning via quantum homomorphic encryption

Quantum-aided Secure Deep Neural Network Inference on Real Quantum Computers

Federated Learning with Quantum Computing and Fully Homomorphic Encryption: A Novel Computing Paradigm Shift in Privacy-Preserving ML

CDQKL: Consensus-Based Distributed Quantum Kernel Learning

Distributed and Secure Kernel-Based Quantum Machine Learning

Non-Unitary Quantum Machine Learning

Drastic Circuit Depth Reductions with Preserved Adversarial Robustness by Approximate Encoding for Quantum Machine Learning

Quantum Private Distributed Learning Through Blind Quantum Computing

A Quantization-based Technique for Privacy Preserving Distributed Learning

Harnessing Inherent Noises for Privacy Preservation in Quantum Machine Learning

Privacy-preserving Decentralized Deep Learning with Multiparty Homomorphic Encryption

Distributed additive encryption and quantization for privacy preserving federated deep learning

Quantum federated learning through blind quantum computing

Quantum Machine Learning: Performance and Security Implications in Real-World Applications

Enhancing Quantum Security over Federated Learning via Post-Quantum Cryptography

Adversarial Robustness Guarantees for Quantum Classifiers

Leveraging Data Locality in Quantum Convolutional Classifiers

Delegated quantum neural networks for encrypted data

Prospects of Privacy Advantage in Quantum Machine Learning