| 引用本文： |
-
王志伟,赵路坦,程佳豪,侯锐,孟丹.基于同态加密的隐私保护神经网络研究综述[J].信息安全学报,已采用 [点击复制]
- wangzhiwei,zhaolutan,chengjiahao,hourui,mengdan.A Survey on Privacy-preserving Neural Networks Based on Homomorphic Encryption[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 深度学习作为一种基于神经网络的机器学习方法，在图像、语音、自然语言处理等领域获得了空前的成功。相较于其他神经网络技术，深度学习模型具有复杂性和可扩展性，能够有效地对大规模数据进行建模，并在实际应用中展现出色的性能。而获得一个高质量的深度学习模型往往需要大量的专家知识和计算资源。随着云计算的广泛普及，云服务器能够为这类解决复杂任务和数据的神经网络技术提供强大算力。在此背景下，“机器学习即服务”的计算模式应运而生。然而，一系列数据安全问题也随之而来。例如，客户端将本地未加密数据上传至云服务器，将失去数据的访问控制权，导致潜在的隐私泄露风险。因此，越来越多的隐私保护法规严格禁止企业或组织收集、分发以及使用用户数据。同态加密作为一种有前途的隐私保护技术，提供了直接在加密数据上计算的能力。基于同态加密实现隐私保护神经网络能够允许不受信任的第三方在不解密的情况下处理数据，从而保护客户端的敏感信息不被泄露。因此，如何高效实现基于同态加密的隐私保护神经网络已成为重要的热点研究方向。本文通过调研现有研究工作，深入分析了基于同态加密实现神经网络推理和训练所面临的问题和挑战，归纳总结了神经网络各层计算与同态加密的结合方法以及相关优化实现。最后，本文总结了基于同态加密的隐私保护神经网络应用存在的关键挑战和未来研究方向。 |
| 关键词: 同态加密 机器学习 隐私保护神经网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2025.04.01 |
| 投稿时间：2023-05-06修订日期：2023-07-05 |
| 基金项目:国家杰出青年科学基金、国家自然科学基金青年科学基金项目 |
|
| A Survey on Privacy-preserving Neural Networks Based on Homomorphic Encryption |
|
wangzhiwei, zhaolutan, chengjiahao, hourui, mengdan
|
| (Institute of Information Engineering, CAS) |
| Abstract: |
| Deep learning, as a machine learning method based on neural networks, has achieved unparalleled success in various fields, including image processing, speech recognition, and natural language processing. In contrast to other neural network technologies, deep learning models are known for their complexity and scalability, enabling effective modeling of large-scale data and delivering exceptional performance in real-world scenarios. However, it"s important to note that obtaining high-quality deep learning models often demands substantial expertise and computational resources. Fortunately, with the widespread adoption of cloud computing, cloud servers offer robust computational capabilities to facilitate the utilization of such neural network technologies in tackling complex tasks and processing data. In this context, the "Machine Learning as a Service" came into being. However, a series of data security issues also follow. For example, if the client uploads local unencrypted data to the cloud server, the data access control will be lost, resulting in potential privacy leakage risks. Therefore, a growing number of privacy protection acts strictly prohibit businesses or organizations from collecting, distributing, and using user data. Homomorphic encryption, as a promising privacy-preserving technique, provides the ability to compute directly on encrypted data. Homomorphic encryption-based privacy-preserving neural network allows an untrusted third party to process the data without decrypting it, thus protecting the client"s sensitive information from being leaked. Therefore, how to efficiently implement homomorphic encryption-based privacy protection neural network has become an important hot research direction. Through the investigation of existing research work, this paper deeply analyzes the problems and challenges faced in the neural network reference and training implementations based on homomorphic encryption, and summarizes the combination method of neural network and homomorphic encryption and related optimization implementation. Finally, this paper summarizes the key challenges and future research directions in privacy-preserving neural network applications based on homomorphic encryption. |
| Key words: homomorphic encryption, neural networks, privacy-preserving machine learning |
