Efficient CNN Building Blocks for Encrypted Data Nayna Jain1,4, Karthik Nandakumar2, Nalini Ratha3, Sharath Pankanti5, Uttam Kumar 1 1 Center for Data Sciences, International Institute of Information Technology, Bangalore 2 Mohamed Bin Zayed University of Artificial Intelligence 3 University at Buffalo, SUNY 4 IBM Systems 5 Microsoft
[email protected],
[email protected],
[email protected]/
[email protected],
[email protected],
[email protected] Abstract Model Owner Model Architecture 푴 Machine learning on encrypted data can address the concerns Homomorphically Encrypted Model Model Parameters 퐸(휽) related to privacy and legality of sharing sensitive data with Encryption untrustworthy service providers, while leveraging their re- Parameters 휽 sources to facilitate extraction of valuable insights from oth- End-User Public Key Homomorphically Encrypted Cloud erwise non-shareable data. Fully Homomorphic Encryption Test Data {퐸(퐱 )}푇 Test Data 푖 푖=1 FHE Computations Service 푇 Encryption (FHE) is a promising technique to enable machine learning {퐱푖}푖=1 퐸 y푖 = 푴(퐸 x푖 , 퐸(휽)) Provider and inferencing while providing strict guarantees against in- Inference 푇 Decryption formation leakage. Since deep convolutional neural networks {푦푖}푖=1 Homomorphically Encrypted Inference Results {퐸(y )}푇 (CNNs) have become the machine learning tool of choice Private Key 푖 푖=1 in several applications, several attempts have been made to harness CNNs to extract insights from encrypted data. How- ever, existing works focus only on ensuring data security Figure 1: In a conventional Machine Learning as a Ser- and ignore security of model parameters. They also report vice (MLaaS) scenario, both the data and model parameters high level implementations without providing rigorous anal- are unencrypted.