Efficient machine learning over encrypted data with non-interactive communication

Efficient Secure Building Blocks With Application to Privacy Preserving Machine Learning Algorithms

“…Further improvements of [16] were seen in [26][27][28], which made FHE suitable for practical applications, hence resulted in the development of actual libraries such as IBM's HElib [31] based on the BGV scheme from [26] and Microsoft's SEAL [29] based on both the SWHE FV presented scheme in [27] and the BGV scheme. Researches realized the strengths of such schemes and immediately started to utilize them for privacy preserving ML algorithms such as deep learning and neural networks [3,17], support vector machines (SVM) [20], random forests and decision trees [44], kNN [18], Naïve Bayes [20][21][22][23][24][25], etc.…”

“…While doing so we take into consideration only horizontally partitioned data and schemes that use cryptographic tools. Despite its simplicity, NB still gives solid accuracy results in areas such health, spam detection, document classification, etc., and therefore is one of the top used classification algorithms [20][21][22][23]. The first scheme to address the issue of privacy preserving (PP) NB training is due to Kantarcıoğlu et.al.…”

“…Without addressing the training part, Park et. al [22] also dealt with the classification problem, by proposing a one-round scheme to overcome the multi-round communication overhead of [15]; but in the process it uses heavy inefficient FHE computations. Li et.al.…”

CryptoNN: Training Neural Networks over Encrypted Data

“…The scheme includes three parts: pre-process-encryption function, pre-process-key-derivative function, and securecomputation function. As a client, it just needs to encrypt the matrix using the pre-process-encrypt function and send out the ciphertext (lines [14][15][16][17][18][19][20][21]. For the server, it first needs to decide the specific function f ∈ F, and then request/prepare the function derived key sk f from the authority (lines 25-30).…”

“…To tackle the challenge of training a model over encrypted data in a simpler manner, in this paper, we propose a novel [3], [9], [10], [11], [12], [13], [14], [15] • • Covers All Homomorphic Encryption (HE) [2] •…”