Efficient Neural Network Inference Protocol Based on Secure Two-Party Computation

doi:10.3969/j.issn.1671-1122.2023.07.003

Abstract

Abstract:

Despite the rapid growth of Machine Learning as a Service(MLaaS) in recent years, there are still many performance and security issues in real-world applications, with the risk of leakage of user data and enterprise neural network model parameters. There are currently some machine learning schemes with privacy protection, but there are problems with low computational efficiency and high communication overhead. To address the above problems, the paper proposed an efficient neural network inference protocol based on secure two-party computation, where the linear layer used secret sharing to protect the privacy of the input data and the nonlinear layer used a low communication overhead comparison function based on oblivious transfer to compute the activation function. The experimental results show that the protocol is at least 23% more efficient and reduces communication overhead by at least 51% on two benchmark datasets compared to existing solutions.

Key words: machine learning, secure two-party computation, neural network, secret sharing, oblivious transfer

CLC Number:

TP309

XU Chungen, XUE Shaokang, XU Lei, ZHANG Pan. Efficient Neural Network Inference Protocol Based on Secure Two-Party Computation[J]. Netinfo Security, 2023, 23(7): 22-30.

Figures/Tables 9

References 24

[1]	RATHOR K, PATIL K, TARUN M S, et al. A Novel and Efficient Method to Detect the Face Coverings to Ensure the Safety Using Comparison Analysis[C]// IEEE. The International Conference on Edge Computing and Applications. New York: IEEE, 2022: 1664-1667.
[2]	BHANGALE K, KOTHANDARAMAN M. Speech Emotion Recognition Based on Multiple Acoustic Features and Deep Convolutional Neural Network[J]. Electronics, 2023, 12(4): 839-855. doi: 10.3390/electronics12040839 URL
[3]	ZHANG Ge, ZHEN Peng, YAN Chaokun, et al. A Novel Liver Cancer Diagnosis Method Based on Patient Similarity Network and DenseGCN[EB/OL]. (2022-12-01) [2023-01-15]. https://www.nature.com/arti-cles/s41598-022-10441-3.
[4]	ZIOVIRIS G, KOLOMVATSOS K, STAMOULIS G. Credit Card Fraud Detection Using a Deep Learning Multistage Model[J]. The Journal of Supercomputing, 2022, 78(12): 14571-14596. doi: 10.1007/s11227-022-04465-9
[5]	YAO A C. Protocols for Secure Computations[C]// IEEE. 23rd Annual Symposium on Foundations of Computer Science. New York: IEEE, 1982: 160-164.
[6]	BOGETOFT P, CHRISTENSEN D L, DAMGARD I, et al. Secure Multiparty Computation Goes Live[C]// Springer. 13th International Conference on Financial Cryptography and Data Security. Heidelberg: Springer, 2009: 325-343.
[7]	MA J P, CHOW S. Secure-Computation-Friendly Private Set Intersection from Oblivious Compact Graph Evaluation[C]// ACM. 2022 ACM on Asia Conference on Computer and Communications Security. New York: ACM, 2022: 1086-1097.
[8]	SPERLING L, RATHA N, ROSS A, et al. HEFT: Homomorphically Encrypted Fusion of Biometric Templates[C]// IEEE. 2022 IEEE International Joint Conference on Biometrics. New York: IEEE, 2022: 1-10.
[9]	MOHASSEL P, ZHANG Yupeng. Secureml: A System for Scalable Privacy-Preserving Machine Learning[C]// IEEE. 2017 IEEE Symposium on Security and Privacy. New York: IEEE, 2017: 19-38.
[10]	RIAZI M S, WEINERT C, TKACHENKO O, et al. Chameleon: A Hybrid Secure Computation Framework for Machine Learning Applications[C]// ACM. The 2018 on Asia Conference on Computer and Communications Security. New York: ACM, 2018: 707-721.
[11]	JUVEKAR C, VAIKUNTANATHAN V, CHANDRAKASAN A. Gazelle: A Low Latency Framework for Secure Neural Network Inference[C]// USENIX. 27th USENIX Security Symposium. Baltimore: USENIX Association, 2018: 1651-1669.
[12]	GARAY J A, SCHOENMAKERS B, VILLEGAS J. Practical and Secure Solutions for Integer Comparison[C]// Springer. 10th International Conference on Practice and Theory in Public-Key Cryptography. Heidelberg: Springer, 2007: 330-342.
[13]	RABIN M O. How to Exchange Secrets with Oblivious Transfer[EB/OL]. [2023-01-15]. https://e-print.iacr.org/2005/187.pdf.
[14]	ASHAROV G, LINDELL Y, SCHNEIDER T, et al. More Efficient Oblivious Transfer and Extensions for Faster Secure Computation[C]// ACM. 2013 ACM SIGSAC Conference on Computer & Communications Security. New York: ACM, 2013: 535-548.
[15]	SHAMIR A. How to Share a Secret[C]// ACM. Communications of the ACM. New York: ACM, 1979: 612-613.
[16]	BEAVER D. Precomputing Oblivious Transfer[C]// Springer. 15th Annual International Cryptology Conference. Heidelberg: Springer, 1995: 97-109.
[17]	DAMGARD I, PASTRO V, SMART N, et al. Multiparty Computation from Somewhat Homomorphic Encryption[C]// Springer. 32nd Annual Cryptology Conference. Heidelberg: Springer, 2012: 643-662.
[18]	YANG Kang, WENG Chenkai, LAN Xiao, et al. Ferret: Fast Extension for Correlated OT with Small Communication[C]// ACM. 2020 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2020: 1607-1626.
[19]	RATHEE D, RATHEE M, KUMAR N, et al. Cryptflow2: Practical 2--Party Secure Inference[C]// ACM. The 2020 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2020: 325-342.
[20]	GOLDWASSER S, MICALI S. Probabilistic Encryption[J]. Journal of Computer and System Sciences, 1984, 28(2): 270-299. doi: 10.1016/0022-0000(84)90070-9 URL
[21]	LIU Jian, JUUTI M, LU Yu, et al. Oblivious Neural Network Predictions via Minionn Transformations[C]// ACM. 2017 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2017: 619-631.
[22]	LIU Xiaoning, ZHENG Yifeng, YUAN Xingliang, et al. Medisc: Towards Secure and Lightweight Deep Learning as a Medical Diagnostic Service[C]// Springer. European Symposium on Research in Computer Security. Heidelberg: Springer, 2021: 519-541.
[23]	WAGH S, TOPLE S, BENHAMOUDA F, et al. Falcon: Honest-Majority Maliciously Secure Framework for Private Deep Learning[C]// Springer. Proceedings on Privacy Enhancing Technologies. Heidelberg: Springer, 2021: 188-208.
[24]	MISHRA M, LEHMKUHL R, SRINIVASAN A, et al. Delphi: A Cryptographic Inference Service for Neural Networks[C]// USENIX. 29th USENIX Security Symposium. California: USENIX Association, 2020: 2505-2522.

模型	时间开销/s	通信开销/MB	准确度	明文上的准确度
M₁	0.501	2.533	98.03%	99.70%
M₂	0.493	2.534	99.09%	99.70%
M₃	120.435	510.043	90.33%	96.53%

模型	离线阶段		线性层		非线性层
模型	时间/s	通信开销/MB	时间/s	通信开销/MB	时间/s	通信开销/MB
M₁	0.102	1.260	0.096	1.003	0.303	0.271
M₂	0.103	1.260	0.099	1.003	0.293	0.274
M₃	14.012	251.875	15.335	249.535	91.113	8.509

M₁	时间/s	M₂	时间/s	M₃	时间/s
文献[21]方案	1.04	文献[21]方案	9.32	文献[21]方案	544
文献[9]方案	4.88	文献[11]方案	8.96	文献[10]方案	245
文献[10]方案	2.24	文献[22]方案	5.63	文献[22]方案	147
文献[22]方案	0.64	文献[23]方案	8.45	文献[24]方案	153
—	—	—	—	文献[23]方案	183
本文协议	0.50	本文协议	0.49	本文协议	120

M₁	通信开销 /MB	M₂	通信开销/MB	M₃	通信开销/MB
文献[21]方案	15.83	文献[21]方案	657.51	文献[21]方案	9272
文献[9]方案	13.31	文献[11]方案	70.00	文献[10]方案	2650
文献[10]方案	10.47	文献[22]方案	5.16	文献[22]方案	489
文献[22]方案	0.92	文献[23]方案	62.12	文献[24]方案	5100
—	—	—	—	文献[23]方案	1278
本文协议	2.53	本文协议	2.53	本文协议	510

M₁	准确度	M₂	准确度	M₃	准确度
文献[21]方案	97.60%	文献[21]方案	99.00%	文献[21]方案	81.61%
文献[9]方案	93.10%	文献[11]方案	98.91%	文献[10]方案	81.61%
文献[10]方案	99.00%	文献[22]方案	99.00%	文献[22]方案	81.00%
文献[22]方案	98.00%	文献[23]方案	99.15%	文献[24]方案	87.30%
—	—	—	—	文献[23]方案	78.02%
本文协议	98.03%	本文协议	99.09%	本文协议	90.33%