1.Key Laboratory of Aerospace Information Security and Trusted Computing,Ministry of Education,School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,Hubei,China
2.Collaborative Innovation Center of Geospatial Technology,Wuhan 430079,Hubei,China
3.Rizhao Institute of Information Technology,Wuhan University,Rizhao 276800,Shandong,China
Prediction as a Service has garnered significant attention recently. Unfortunately, the outsourced prediction paradigm raises various security issues. To solve the privacy issues of clients and model providers, as well as the lack of verifiability of the cloud servers, this paper proposes a novel system that provides both privacy protection and verifiability for outsourcing deep neural network prediction. Firstly, this paper designs a proprietary interactive proof protocol for deep neural network models based on the sum-check protocols, which can efficiently verify the correctness of the results returned by the cloud servers. Secondly, this paper ensures privacy protection for outsourcing deep neural network predictions by employing additive secret sharing and variants of the D-A multiplication protocol. Leveraging the dual server system model without collusion, the paper utilizes additive secret⁃sharing technology to achieve the privacy protection of the client and the model provider. Then, this paper implements the multiplication of additivity shares based on variants of the D-A protocol. Finally, this paper combines the newly designed interactive proof protocols with variants of the D-A multiplication protocol to design new secure protocols for model prediction. These new protocols successfully achieve privacy protection and verifiability in the deep neural network model prediction outsourcing paradigm. The experimental evaluations conducted on the MNIST and CIFAR-10 datasets demonstrate the practical performance of our proposed scheme, achieving 1.6× to 247.0× improved latency and 1.3× to 27.7× enhanced communication compared to most state-of-the-art solutions.
图灵奖获得者姚期智教授在文献[21]中首次提出了安全多方计算的概念,开启了安全多方计算研究的历程。安全多方计算协议的主要目标是实现由多个参与方协同完成同一个计算任务,同时保证在这个过程中不泄露各参与方的任何私密数据。伴随安全多方计算技术的发展,基于安全多方计算协议的各种隐私保护预测研究也获得了进一步的发展。文献[7]基于秘密共享(secret sharing,SS)[15]和混淆电路[22]提出了MiniONN方案,它是第一种将现有神经网络转换为具有合理效率的支持隐私保护预测的不经意神经网络的方法。与以往的工作不同,MiniONN方案不需要改变模型的训练方式,为神经网络预测模型中常用的操作设计了不经意协议,并采用单指令-多数据技术来提高效率。MiniONN方案在响应延迟和消息大小方面均优于之前的方案。文献[23]提出了Chameleon,它是一种用于安全函数评估(secure function evaluation, SFE)的新框架,该框架使双方能够共同计算一个函数而无需泄露其私人输入。Chameleon框架将通用SFE协议的最优点与基于ASS的协议相结合。几乎所有开销大的加密操作都在离线阶段进行了预计算,大大减少了Chameleon框架的通信开销。文献[24]设计了GAZELLE方案,用于安全神经网络预测,它将同态加密和混淆电路技术相结合,可扩展性好,延迟低。GAZELLE方案的贡献可以总结为以下三点:首先,它设计了GAZELLE同态加密库,该库为单指令-多数据形式的相加、相乘和密文排列等基本同态操作提供了快速算法;其次,它实现了GAZELLE同态线性代数内核,将神经网络层映射到优化的同态矩阵向量乘法和卷积例程;第三,它设计了优化的加密交换协议,在同态和混淆电路编码之间无缝转换,实现了完整的神经网络预测。文献[25]的DELPHI是一个用于现实世界中具有神经网络结构的安全预测系统,允许双方在不泄露任何一方数据的情况下执行神经网络预测。DELPHI方案设计了一个混合密码协议,该协议在通信和计算开销方面都比以前的工作有所下降;DELPHI方案开发了一个自动生成神经网络架构配置的规划器,该规划器很好地实现了混合协议的性能和准确率之间的权衡。
WANGW, WANGS, GAOJ, et al. Rafiki: Machine learning as an analytics service system [J]. Proceedings of the VLDB Endowment, 2018, 12(2): 128-140. DOI: 10.14778/3282495.3282499 .
[2]
AZURE. Machine Learning Studio [EB/OL]. [2022-01-10].
[3]
AWS. Machine Learning on Aws [EB/OL]. [2022-01-13].
DOWLINN, GILAD-BACHRACHR, LAINEK, et al. CryptoNets: Applying neural networks to encrypted data with high throughput and accuracy [C]// Proceedings of the 33rd International Conference on Machine Learning. New York: IEEE, 2016: 201-210. DOI: 10.1109/jproc.2016.2622218 .
[7]
LIUJ, JUUTIM, LUY, et al. Oblivious neural network predictions via MiniONN transformations[C]//Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2017: 619-631. DOI: 10.1145/3133956.3134056 .
[8]
LIM H, YANY J, WANGQ, et al. Secure prediction of neural network in the cloud[J]. IEEE Network, 2021, 35(1): 251-257. DOI: 10.1109/MNET.011.2000293 .
MONROSEF, WYCKOFFP, RUBINA D. Distributed Execution with Remote Audit[DB/OL].[2022-10-23]
[13]
GOLDWASSERS, KALAIY T, ROTHBLUMG N. Delegating computation: Interactive proofs for muggles[J]. Journal of the ACM, 2015,62(4):Article No. 27. DOI: 10.1145/2699436 .
[14]
CORMODEG, MITZENMACHERM, THALERJ. Practical verified computation with streaming interactive proofs[C]//Proceedings of the 3rd Innovations in Theoretical Computer Science Conference. New York: ACM, 2012: 90-112. DOI: 10.1145/2090236.2090245 .
[15]
MIGNOTTEM. How to share a secret [C]// Proceedings of the Workshop on Cryptography. Berlin: Springer, 1982: 371-375.
[16]
GHODSIZ, GUT Y, GARGS. SafetyNets: Verifiable execution of deep neural networks on an untrusted cloud[EB/OL]. 2017: arXiv: 1706.10268.
[17]
GENTRYC, HALEVIS, SMARTN P. Fully homomorphic encryption with polylog overhead[C]//Annual International Conference on the Theory and Applications of Cryptographic Techniques. Berlin: Springer, 2012: 465-482. DOI:10.1007/978-3-642-29011-4_28 .
[18]
BOSJ W, LAUTERK, LOFTUSJ, et al. Improved security for a ring-based fully homomorphic encryption scheme[C]//IMA International Conference on Cryptography and Coding. Berlin: Springer, 2013: 45-64. DOI:10.1007/978-3-642-45239-0_4 .
[19]
SANYALA, KUSNERM J, GASCÓNA, et al. TAPAS: Tricks to accelerate (encrypted) prediction as a service[EB/OL]. 2018: arXiv: 1806.03461.
[20]
BRUTZKUSA, GILAD-BACHRACHR, ELISHAO. Low latency privacy preserving inference [C]// Proceedings of the 36th International Conference on Machine Learning. Long Beach: PMLR, 2019: 812-821.
[21]
YAOA C. Protocols for secure computations[C]//23rd Annual Symposium on Foundations of Computer Science (SFCS 1982). New York: IEEE Press, 2008: 160-164. DOI: 10.1109/SFCS.1982.38 .
[22]
YAOA C C. How to generate and exchange secrets[C]//27th Annual Symposium on Foundations of Computer Science (SFCS 1986). New York: IEEE Press, 2008: 162-167. DOI: 10.1109/SFCS.1986.25 .
[23]
RIAZIM S, WEINERTC, TKACHENKOO, et al. Chameleon: A hybrid secure computation framework for machine learning applications[C]//Proceedings of the 2018 on Asia Conference on Computer and Communications Security. New York: ACM, 2018: 707-721. DOI: 10.1145/3196494.3196522 .
[24]
JUVEKARC, VAIKUNTANATHANV, CHANDRAKASANA. Gazelle: A low latency framework for secure neural network inference[EB/OL]. 2018: arXiv: 1801.05507.
[25]
PRATYUSHM, RyanL, AKSHAYARAMS, et al. DELPHI: A cryptographic inference service for neural net-works [C]// Proceedings of the 29th USENIX Security Symposium. Baltimore: USENIX Association, 2020: 2505-2522.
[26]
LEHMKUHLR, MISHRAP, SRINIVASANA, et al. Muse: Secure inference resilient to malicious clients [C]// Proceedings of the 30th USENIX Security Symposium. Baltimore: USENIX Association, 2021: 2201-2218.
[27]
CHANDRANN, GUPTAD, OBBATTUS L B, et al. SIMC: ML inference secure against malicious clients at semi-honest cost [C]// Proceedings of the 31th USENIX Security Symposium. Boston: USENIX Association, 2022: 1361-1378.
[28]
WALFISHM, BLUMBERGA J. Verifying computations without reexecuting them[J]. Communications of the ACM, 2015, 58(2): 74-84. DOI: 10.1145/2641562 .
[29]
LUNDC, FORTNOWL, KARLOFFH, et al. Algebraic methods for interactive proof systems[C]//Proceedings 31st Annual Symposium on Foundations of Computer Science. New York: IEEE Press, 2002: 2-10. DOI: 10.1109/FSCS.1990.89518 .
[30]
CHELLAPILLAK, PURIS, SIMARDP. High Performance Convolutional Neural Networks for Document Processing [EB/OL]. [2006-11-09].
[31]
GAUTIERA, NGUYENQ, HEINM. Globally optimal training of generalized polynomial neural networks with nonlinear spectral methods[EB/OL]. 2016: arXiv: 1610.09300.
[32]
LIVNIR, SHALEV-SHWARTZS, SHAMIRO. On the computational efficiency of training neural networks[EB/OL]. 2014: arXiv: 1410.1141.
[33]
DUW L, ATALLAHM. Protocols for secure remote database access with approximate matching [J]. E⁃Commerce Security and Privacy, 2001, 2: 87-111. DOI: 10.1007/978-1-4615-1467-1_6 .
[34]
VU V, SETTYS, BLUMBERGA J, et al. A hybrid architecture for interactive verifiable computation[C]//2013 IEEE Symposium on Security and Privacy. New York: IEEE Press, 2013: 223-237. DOI: 10.1109/SP.2013.48 .
[35]
CANETTIR. Universally composable security: A new paradigm for cryptographic protocols[C]//Proceedings 42nd IEEE Symposium on Foundations of Computer Science. New York: IEEE Press, 2002: 136-145. DOI: 10.1109/SFCS.2001.959888 .
[36]
ELGOHARYA, SOBHT S, ZAKIM. Design of an enhancement for SSL/TLS protocols[J]. Computers & Security, 2006, 25(4): 297-306. DOI: 10.1016/j.cose.2006.02.007 .
[37]
SHOKRIR, STRONATIM, SONGC Z, et al. Membership inference attacks against machine learning models[C]//2017 IEEE Symposium on Security and Privacy (SP). New York: IEEE Press, 2017: 3-18. DOI: 10.1109/SP.2017.41 .
MOHASSELP, ZHANGY P. SecureML: A system for scalable privacy-preserving machine learning[C]//2017 IEEE Symposium on Security and Privacy (SP). New York: IEEE Press, 2017: 19-38. DOI: 10.1109/SP.2017.12 .
[40]
LECUNY, CORTESC, BURGESC J. MNIST Dataset [EB/OL]. [2016-04-21].
LOUQ, JIANGL. SHE: A Fast and Accurate Deep Neural Network for Encrypted Data[EB/OL]. [2022-10-13].
[43]
HANS, MAOH Z, DALLYW J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding[EB/OL]. 2015: arXiv: 1510.00149.