現今知名公有雲平台對於個人資料委外於雲端的保護僅限於資料傳輸與存放時的加密保護，不提供使用資料進行計算時的保護，以及對於進行資料分析所使用的機器學習軟體也不提供保護。因此在公有雲平台上無法落實安全式機器學習即服務的軟體與資料共同保護。本研究提出「機器學習即服務」軟體模組，在資料加密以及軟體混淆的共同保護下，來完成資料分析時的正確分類與預測。本研究將使用Kaggle上的“Titanic: Machine Learning from Disaster”資料集，以明文及明碼的方式訓練出最佳化模型，透過Indistinguishability Obfuscation（iO）的Graded Encoding Schemes（GES）技術將資料分析所使用的Support Vector Machine（SVM）二元分類函式及測試資料進行混淆達到程式及資料共同保護，搭配運用5GenCrypto套件進行，來完成進行安全式機器學習於公有雲平台，並具體提出本方法的量化與質化的運算觀察結果。

Nowadays, the protection of personal data on some famous public cloud platforms is applicable only when the data is in transmission or at rest by encryption. It does not protect the data in use, and the machine learning programs for data analysis. Therefore, it cannot protect both program and data for secure Machine Learning as a Service（MLaaS）. This research proposed a MLaaS program model which is able to make correct classification and prediction on data analysis with the protection on both data encryption and program obfuscation. This research used the dataset “Titanic: Machine Learning from Disaster” on Kaggle, and the plaintext to train the best model. Then, we use the Graded Encoding Schemes（GES） method of Indistinguishability Obfuscation（iO）to obfuscate the SVM binary classification hyperplane and test data to ensure both program and data protection. We use 5Gen Crypto package to execute secure machine learning on public cloud platform, and concluding the calculation results of quantization and quality by this method.

[1] Chandramouli, R., et al., Cryptographic Key Management Issues & Challenges in Cloud Services. NISTIR 7956, NIST, U. S. Department of Commerce, 2013.
[2] Damgard, I., et al., Secure Key Management in the Cloud. IMA CC 2013, 2013.
[3] Gentry, C., Computing on the Edge of Chaos: Structure and Randomness in Encrypted Computation. Proc. of the Int. Congress of Mathematicians, Seoul, 2014.
[4] Garg, S. et al., Candidate Indistinguishability Obfuscation and Functional Encryption for All Circuits. FOCS13, pp. 40-49, 2013.
[5] Barrington, A. D., Bounded-Width Polynomial-Size Branching Programs Recognize Exactly Those Language in NC1. Journal of Computer and System Science 38, pp. 150-164, 1989.
[6] Barak, B., Hopes, Fears, and Software Obfuscation. CACM, 59(3), March, 2016.
[7] Garg, S., et al., Hiding Secrets in Software: A Cryptographic Approach to Program Obfuscation. CACM, 59(5), May 2016.
[8] Lewi, K., et al., 5Gen: A Framework for Prototyping Applications Using Multilinear Maps and Matrix Branching Programs. CCS’16, 2016.
[9] Collberg, C. and Nagra, J., Surreptitious Software: Obfuscation, Watermarking,
and Tamerproofing for Software Protection. Wiley, 2009
[10] Horváth, M., Survey on Cryptographic Obfuscation. Cryptology ePrint Archive, Report, 2015/412
[11] Barak, B., et al. On the (Im)possibility of Obfuscating Programs. Journal of
the ACM, 59(2),Apr. 2012.
[12] Sauerhoff, M., et al. Relating branching program size and formula size over the full binary basis. STACS 99: 16th Annual Sysmposium on Theoretical Aspects of Computer Science, volume 1563 of Lecure Notes in Computer Science, pages 57-67, Trier, Gemery, Mar. 4-6 1999.
[13] Apon, D., et al., Implementing Cryptographic Program Obfuscation. ePrint Archive, Report, 2014/779
[14] Garg, S. et al., Candidate multilinear maps from ideal lattices. EUROCRYPT 2013, LNCS 7881, pp. 1–17.
[15] Coron, J. S. et al., Practical multilinear maps over the integers. CRYPTO 2013, LNCS 8042, pp. 476–493.
[16] Cortes, C. and Vapnik, V., Support-Vector Networks. Machine Learning, pp. 273-297, 1995.
[17] Gentry, G., Fully Homomorphic Encryption Using Ideal Lattices. STOC’09, 2009.
[18] Fan, J. and F. Vercauteren, Somewhat Practically Fully Homomorphic Encryption. ICAR Cryptology ePrint archive, 2012.
[19] Bost, R., Machine learning classification over encrypted data. NDSS’15, Feb. 2015.
[20] Graepel, T., et al., ML Confidential: Machine Learning on Encrypted Data. Information Security and Cryptology – ICISC, LNCS, Springer, 2012.
[21] Collberg, C. et al.,. A Taxonomy of Obfuscating Transformations. Computer Science Technical Reports 148, 1997.