Privacypreserving logistic regression training
 279 Downloads
Abstract
Background
Logistic regression is a popular technique used in machine learning to construct classification models. Since the construction of such models is based on computing with large datasets, it is an appealing idea to outsource this computation to a cloud service. The privacysensitive nature of the input data requires appropriate privacy preserving measures before outsourcing it. Homomorphic encryption enables one to compute on encrypted data directly, without decryption and can be used to mitigate the privacy concerns raised by using a cloud service.
Methods
In this paper, we propose an algorithm (and its implementation) to train a logistic regression model on a homomorphically encrypted dataset. The core of our algorithm consists of a new iterative method that can be seen as a simplified form of the fixed Hessian method, but with a much lower multiplicative complexity.
Results
We test the new method on two interesting real life applications: the first application is in medicine and constructs a model to predict the probability for a patient to have cancer, given genomic data as input; the second application is in finance and the model predicts the probability of a credit card transaction to be fraudulent. The method produces accurate results for both applications, comparable to running standard algorithms on plaintext data.
Conclusions
This article introduces a new simple iterative algorithm to train a logistic regression model that is tailored to be applied on a homomorphically encrypted dataset. This algorithm can be used as a privacypreserving technique to build a binary classification model and can be applied in a wide range of problems that can be modelled with logistic regression. Our implementation results show that our method can handle the large datasets used in logistic regression training.
Keywords
Homomorphic encryption Logistic regression Privacy Fixed HessianAbbreviations
 AUC
Area under the curve
 CRT
Chinese remainder theorem
 FN
False negative
 FP
False positive
 FPR
False positive rate
 FV
Fan vercauteren
 KeyGen
Key generation
 LWE
Learning with error
 RNS
Residue number system
 ROC
Receiver operating characteristic
 SFH
Simplified fixed Hessian
 SIMD
Single instruction multiple data
 TN
True negative
 TP
True positive
 TPR
True positive rate
 wNIBNAF
w non integral base non adjacent form
Background
Introduction
Logistic regression is a popular technique used in machine learning to solve binary classification problems. It starts with a training phase during which one computes a model for prediction based on previously gathered values for predictor variables (called covariates) and corresponding outcomes. The training phase is followed by a testing phase that assesses the accuracy of the model. To this end, the dataset is split into data for training and data for validation. This validation is done by evaluating the model in the given covariates and comparing the output with the known outcome. When the classification of the model equals the outcome for most of the test data, the model is considered to be valuable and it can be used to predict the probability of an outcome by simply evaluating the model for new measurements of the covariates.
Logistic regression is popular because it provides a simple and powerful method to solve a wide range of problems. In medicine, logistic regression is used to predict the risk of developing a certain disease based on observed characteristics of the patient. In politics, it is used to predict the voting behaviour of a person based on personal data such as age, income, sex, state of residence, previous votes. In finance, logistic regression is used to predict the likelihood of a homeowner defaulting on a mortgage or a credit card transaction being fraudulent.
As all machine learning tools, logistic regression needs sufficient training data to construct a useful model. As the above examples show, the values for the covariates and the corresponding outcomes are typically highly sensitive, which implies that the owners of this data (either people or companies) are reluctant to have their data included in the training set. In this paper, we solve this problem by describing a method for privacy preserving logistic regression training using somewhat homomorphic encryption. Homomorphic encryption enables computations on encrypted data without needing to decrypt the data first. As such, our method can be used to send encrypted data to a central server, which will then perform logistic regression training on this encrypted input data. This also allows to combine data from different data owners since the server will learn nothing about the underlying data.
Related work
Private logistic regression with the aid of homomorphic encryption has already been considered in [1, 2], but in a rather limited form: both papers assume that the logistic model has already been trained and is publicly available. This publicly known model is then evaluated on homomorphically encrypted data in order to perform classification of this data without compromising the privacy of the patients. Our work complements these works by executing the training phase for the logistic regression model in a privacypreserving manner. This is a much more challenging problem than the classification of new data, since this requires the application of an iterative method and a solution for the nonlinearity in the minimization function.
Aono et al. [3] also explored secure logistic regression via homomorphic encryption. However, they shift the computations that are challenging to perform homomorphically to trusted data sources and a trusted client. Consequently, in their solution the data sources need to compute some intermediate values, which they subsequently encrypt and send to the computation server. This allows them to only use an additively homomorphic encryption scheme to perform the second, easier, part of the training process. Finally, they require a trusted client to perform a decryption of the computed coefficients and use these coefficients to construct the cost function for which the trusted client needs to determine the minimum in plaintext space. Their technique is based on a polynomial approximation of the logarithmic function in the cost function and the trusted client applies the gradient descent algorithm as iterative method to perform the minimization of the cost function resulting from the homomorphic computations. Our method does not require the data owners to perform any computations (bar the encryption of their data) and determines the model parameters by executing the minimization directly on encrypted data. Again this is a much more challenging problem.
In [4] Xie et al. construct PrivLogit which performs logistic regression in a privacypreserving but distributed manner. As before, they require the data owners to perform computations on their data before encryption to compute parts of a matrix used in the logistic regression. Our solution starts from the encrypted raw dataset, not from values that were precomputed by the centers that collect the data. In our solution all computations that are needed to create the model parameters, are performed homomorphically.
Independently and in parallel with our research, Kim et al. [5] investigated the same problem of performing the training phase of logistic regression in the encrypted domain. Their method uses a different approach than ours: firstly, they use a different minimization method (gradient descent) compared to ours (a simplification of the fixed Hessian method), a different approximation of the sigmoid function and a different homomorphic encryption scheme. Their solution is based on a small adaptation of the input values, which reduces the number of homomorphic multiplications needed in the computation of the model. We assumed the dataset would be already encrypted and therefore adaptations to the input would be impossible. Furthermore, they tested their method on datasets that contain a smaller number of covariates than the datasets used in this article.
Contributions
Our contributions in this paper are as follows: firstly, we develop a method for privacy preserving logistic training using homomorphic encryption that consists of a low depth version of the fixed Hessian method. We show that consecutive simplifications result in a practical algorithm, called the simplified fixed Hessian (SFH) method, that at the same time is still accurate enough to be useful. We implemented this algorithm and tested its performance and accuracy on two real life use cases: a medical application predicting the probability of having cancer given genomic data and a financial application predicting the probability that a transaction is fraudulent. Our test results show that in both use cases the model computed is almost as accurate as the model computed by standard logistic regression tools such as the ones present in Matlab.
Technical Background
Logistic regression
where the vector β=(β_{0},…,β_{d}) are the model parameters, y the class label (in our case {±1}) and the vector \(\mathbf {x} = (1, x_{1}, \ldots, x_{d}) \in \mathbb {R}^{d+1}\) the covariates.
Datasets
As mentioned before, we will test our method in the context of two real life use cases, one in genomics and the other in finance.
The genomic dataset was provided by the iDASH competition of 2017 and consists of 1581 records (each corresponding to a patient) consisting of 103 covariates and a class variable indicating whether or not the patient has cancer. The challenge was to devise a logistic regression model to predict the disease given a training data set of at least 200 records and 5 covariates. However, for scalability reasons the solution needed to be able to scale up to 1000 records with 100 covariates. This genomic dataset consists entirely of binary data.
The financial data was provided by an undisclosed bank that provided anonymized data with the goal of predicting fraudulent transactions. Relevant data fields that were selected are: type of transaction, effective amount of the transaction, currency, origin and destination, fees and interests, etc. This data has been subject to preprocessing by firstly representing the nonnumerical values with labels and secondly computing the minimum and maximum for each of the covariates and using these to normalise the data by computing \(\frac {xx_{\text {min}}}{x_{\text {max}}x_{\text {min}}}\). The resulting financial dataset consists of 20,000 records with 32 covariates, containing floating point values between 0 and 1.
The FV scheme
Our solution is based on the somewhat homomorphic encryption scheme of Fan and Vercauteren [6], which can be used to compute a limited number of additions and multiplications on encrypted data. The security of this encryption scheme is based on the hardness of the ring learning with error problem (RLWE) introduced by Lyubashevsky et al. in [7]. The core objects in the FV scheme are elements of the polynomial ring \(R=\mathbb {Z}[X]\slash (f(X))\), where typically one chooses f(X)=X^{D}+1 for D=2^{n} (in our case D=4096). For an integer modulus \(M \in \mathbb {Z}\) we denote with R_{M} the quotient ring R/(MR).
The plaintext space of the FV scheme is the ring R_{t} for t>1 a small integer modulus and the ciphertext space is R_{q}×R_{q} for an integer modulus q≫t. For a∈R_{q}, we denote by [a]_{q} the element in R obtained by applying [·]_{q} to all its coefficients a_{i}, with [a_{i}]_{q}=a_{i} mod q given by a representative in \(\left (\frac {q}{2},\frac {q}{2}\right ]\). The FV scheme uses two probability distributions on R_{q}: one is denoted by χ_{key} and is used to sample the secret key of the scheme, the other is denoted χ_{err} and will be used to sample error polynomials during encryption. The exact security level of the FV scheme is based on these probability distributions, the degree D and the ciphertext modulus q and can be determined using an online tool developed by Albrecht et al. [8].

KeyGen: the private key consists of an element s←χ_{key} and the public key pk=(b,a) is computed as a←R_{q} uniformly at random and b=[−(as+e)]_{q} with e←χ_{err}.

Encrypt(pk, m): given m∈R_{t}, sample error polynomials e_{1},e_{2}∈χ_{err} and u∈χ_{key} and compute c_{0}=Δm+bu+e_{1} and c_{1}=au+e_{2} with Δ=⌊q/t⌋, the largest integer smaller than \(\frac {q}{t}\). The ciphertext is then c=(c_{0},c_{1}).

Decrypt(sk, c): compute \(\tilde {m}=[c_{0}+c_{1} s]_{q}\), divide the coefficients of \(\tilde {m}\) by Δ and round and reduce the result into R_{t}.
Computing the sum of two ciphertexts simply amounts to adding the corresponding polynomials in the ciphertexts. Multiplication, however, requires a bit more work and we refer to [6] for the precise details.
The relation between a ciphertext and the underlying plaintext can be described as [c_{0}+c_{1}s]_{q}=Δm+e, where e is the noise component present in the ciphertext. This also shows that if the noise e grows too large, decryption will no longer result in the original message, and the scheme will no longer be correct. Since the noise present in the resulting ciphertext will grow with each operation we perform homomorphically, it is important to choose parameters that guarantee correctness of the scheme. Knowing the computations that need to be performed up front enables us to estimate the size of the noise in the resulting ciphertext, which permits the selection of suitable parameters.
wNIBNAF
A final step then maps this Laurent polynomial into the plaintext space R_{t} and we refer the reader to [9] for the precise details.
The wNIBNAF encoding is constructed such that the encoding of a number will satisfy two conditions: the encoding has coefficients in the set {−1,0,1} and each set of w consecutive coefficients will have no more than one nonzero coefficient. Both conditions ensure that the encoded numbers are represented by very sparse polynomials with coefficients in the set {−1,0,1}, which can be used to bound the size of the coefficients of the result of computations on these representations. In particular, this encoding results in a smaller plaintext modulus t, which improves the performance of the homomorphic encryption scheme. Since larger values for w increase the sparseness of the encodings and hence reduce the size of t even more, one would like to select the value for w to be as large as possible. However, similar to encryption one has to consider a correctness requirement for the encoding. More specifically, decoding of the final polynomial should result in the correct answer, hence the base b_{w} and consequently also the value of w should be chosen with care.
Methods
Privacy preserving training of the model
NewtonRaphson method
Homomorphic logistic regression
The downside of Newton’s method is that exact evaluation of the Hessian and its inverse are quite expensive in computational terms. In addition, the goal is to estimate the parameters of the logistic regression model in a privacypreserving manner using homomorphic encryption, which will further increase the computational challenges. Therefore, we will adapt the method in order to make it possible to compute it efficiently in the encrypted domain.
Furthermore, they suggest a lower bound specifically for the Hessian of the logistic regression problem, which is defined as \(\bar {H} = \frac {1}{4} X^{T}X\) and demonstrate that this is a good bound. This approximation does not depend on β, consequently it is fixed throughout all iterations and it only needs to be computed once as desired. Since the Hessian is fixed, so is its inverse, which means it only needs to be computed once.
To be able to use this approximation as lower bound for the above fixed Hessian method we need to assure ourselves it satisfies the condition \(H(\boldsymbol {\beta })\geq \tilde {H}\). As mentioned before we already know from [10] that \(H(\boldsymbol {\beta })\geq \frac {1}{4}X^{T}X\), so it is sufficient to show that \(\frac {1}{4}X^{T}X\geq \tilde {H}\), which we now prove more generally.
Lemma 1
Let \(A \!\in \! \mathbb {R}^{n \times n}\) be a symmetric matrix with all entries nonpositive, and let B be the diagonal matrix with diagonal entries \(B_{k,k} \,=\, \sum _{i = 1}^{n} A_{k,i}\) for k = 1,…,n, then A ≥ B.
Proof
Note that by construction of C we have that \(C_{i,i} = \sum _{j\neq i}{C_{ij}}\), and so every eigenvalue λ satisfies λ−C_{i,i}<C_{i,i} for some i. In particular, since C_{i,i}≥0, we conclude that λ≥0 for all eigenvalues λ and thus that A≥B. □
Our approximation \(\tilde {H}\) for the Hessian also simplifies the computation of the inverse of the matrix, since we simply need to invert each diagonal element separately. The inverse will be again computed using the NewtonRaphson method: assume we want to invert the number a, then the function f(x) will be equal to \(\frac {1}{x}a\) and the iteration is given by x_{k+1}=x_{k}(2−ax_{k}). For the NewtonRaphson method to converge, it is important to determine a good start value. Given the value range of the input data and taking into account the dimensions of the training data, we estimate a range of the size of the number we want to invert. This results in an estimation of the order of magnitude of the solution that is expected to be found by the NewtonRaphson algorithm. By choosing the initial value of our NewtonRaphson iteration close to the constructed estimation of the inverse, we can already find an acceptable approximation of the inverse by performing only one iteration of the method.
In the third and final step, we simplify the nonlinearity coming from the sigmoid function. Here, we simply use the Taylor series: extensive experiments with plaintext data showed that approximating σ(y_{i}β^{T}x_{i}) by \(\frac {1}{2}+\frac {y_{i}\boldsymbol {\beta }^{T}\mathbf {x}_{i}}{4}\) is enough to obtain good results.
The combination of the above techniques finally results in our simplified fixed Hessian (SFH) method given in Algorithm 1.
Performance for the financial dataset with 31 covariates and 700 training records and 19,300 testing records
# iterations  AUC SFH 

1  0.9418 
5  0.9436 
10  0.9448 
20  0.9466 
50  0.9517 
100  0.9599 
Results
Accuracy of the SFH method
Comparing actual and predicted classes
Actual class  

1  1  
Predicted  1  True negative (TN)  False negative (FN) 
Class  1  False positive (FP)  True positive (TP) 
By computing the TPR and FPR for varying thresholds 0≤τ≤1, we can construct the receiver operating characteristic curve or ROCcurve. The ROCcurve is constructed by plotting the (FPR,TPR) pairs for each possible value of the threshold τ. In the ideal situation there would exists a point with (FPR,TPR)=(0,1), which would imply that there exists a threshold for which the model classifies all test data correctly.
The area under the ROCcurve or AUCvalue will be used as the main indicator of how well the classifier works. Since our SFH method combines several approximations, we need to verify the accuracy of our model first on unencrypted data and later on encrypted data. For well chosen system parameters, there will be no difference between accuracy for unencrypted vs. encrypted data since all computations on encrypted data are exact.
The first step is performed by comparing our SFH method with the standard logistic regression functionality of Matlab. This is done by applying our method with all its approximations to the plaintext data and comparing the result to the result of the “glmfit” function in Matlab. The function b=glmfit(X,y,distr) returns a vector b of coefficient estimates for a generalized linear model of the responses y on the predictors in X, using distribution distr. Generalized linear models unify various statistical models, such as linear regression, logistic regression and Poisson regression, by allowing the linear model to be related to the response variable via a link function. We use the “binomial” distribution, which corresponds to the “logit” link function and y a binary vector indicating success or failure to compute the parameters of the logistic regression model with “glmfit”.
Implementation details and performance
 1
the security of the somewhat homomorphic FV scheme,
 2
the correctness of the somewhat homomorphic FV scheme,
 3
the correctness of the wNIBNAF encoding.
The security of a given set of system parameters can be estimated using the work of Albrecht, Player and Scott [13] and the open source learning with error (LWE) hardness estimator implemented by Albrecht [8]. This program estimates the security of the LWE problem based on the following three parameters: the degree D of the polynomial ring, the ciphertext modulus q and \(\alpha = \frac {\sqrt {2\pi }\sigma }{q}\) where σ is the standard deviation of the error distribution χ_{err}. The security estimation is based on the best known attacks for the learning with error problem. Our system parameters are chosen to be q=2^{186}, D=4096 and σ=20 (and thus \(\alpha =\frac {\sqrt {2\pi }\sigma }{q}\)) which results in a security of 78 bits.
As explained in the section on the FV scheme, the error in the ciphertext encrypting the result, should be small enough to enable correct decryption. By estimating the infinity norm of the noise we can select parameters that keep this noise under the correctness bound and in particular, we obtain an upper bound t_{max} of the plaintext modulus. Similarly, to ensure correct decoding, the coefficients of the polynomial encoding the result must remain smaller than the size of the plaintext modulus t. This condition results in a lower bound on the plaintext modulus t_{min}.
The parameters defining plaintext encoding
w  t  

Genomic data  (1)  71  5179·5189·5197 
Financial data  (2)  150  2237·2239 
Since we are using the Chinese Remainder Theorem, each record will be encrypted using two (for the financial fraud case) or three (for the genomics case) ciphertexts. As such, a timememory trade off is possible depending on the requirements of the application. One can choose to save computing time by executing the algorithm for the different ciphertexts in parallel; or one can choose to save memory by computing the result for each plaintext space \(R_{t_{i}}\) consecutively and overwriting the intermediate values of the computations in the process.
The time complexity of our SFH method is also easy to estimate, but one has to be careful to perform the operations in a specific order. If one would naively compute the matrix \(\tilde {H}\) by first computing \(\bar {H}\) and subsequently summing each row, the complexity would be O(Nd^{2}). However, the formula of the kth diagonal element of \(\tilde {H}\) is given by \(\frac {1}{4}\sum _{j=1}^{d+1}\left (\sum _{i=1}^{N}{x_{k,i}x_{j,i}}\right)\), which can be rewritten as \(\frac {1}{4} \sum _{i=1}^{N}{x_{k,i}} \left (\sum _{j = 1}^{d+1} {x_{j,i}} \right)\). This formula shows that it is more efficient to first sum all the rows of X and then perform a matrix vector multiplication with complexity O(Nd).
Performance for the genomic dataset with a fixed number of covariates equal to 20
# training records  Computation time  AUC SFH  AUC glmfit 

500  22 min  0.6348  0.6287 
600  26 min  0.6298  0.6362 
800  35 min  0.6452  0.6360 
1000  44 min  0.6561  0.6446 
Performance for the genomic dataset with a fixed number of training records equal to 500 and the number of testing records equal to 1081
# covariates  Computation time  AUC SFH  AUC glmfit 

5  7 min  0.65  0.6324 
10  12 min  0.6545  0.6131 
15  17 min  0.6446  0.6241 
20  22 min  0.6348  0.6272 
Performance for the financial dataset with a fixed number of covariates equal to 31
# training records  Computation time  AUC SFH  AUC glmfit 

700  30 min  0.9416  0.9619 
800  36 min  0.9411  0.9616 
900  40 min  0.9409  0.9619 
1000  45 min  0.9402  0.9668 
Performance for the financial dataset with a fixed number of records equal to 500 and the number of testing records equal to 19,500
# covariates  Computation time  AUC SFH  AUC glmfit 

5  5 min  0.8131  0.8447 
10  8 min  0.9403  0.9409 
15  11 min  0.9327  0.9492 
20  15 min  0.9401  0.9629 
In Tables 4 and 5 we see that often the AUC value of the SFH model is slightly higher than the AUC value of the glmfit model. However, as mentioned before both models perform poorly on this dataset. Since our SFH model contains many approximations we expect it to perform slightly worse than the “glmfit” model. Only slightly worse because Figs. 1 and 2 already showed that the SFH models classifies the data almost as well as the “glmfit” model. This is consistent with the results for the financial dataset shown in Tables 6 and 7, which we consider more relevant than the results of the genomic dataset due to the fact that both models perform better on this dataset.
Discussion
The experiments of this article show promising results for the simple iterative method we propose as an algorithm to compute the logistic regression model. A first natural question is whether this technique is generalizable to other machine learning problems. In [14], Böhning describes how to adapt the lower bound method to make it applicable to multinomial logistic regression, it is likely this adaption will also apply to our SFH technique and hence our SFH technique can most likely also be applied to construct a multinomial logistic regression model. In the case of neural networks we can refer to [15]; in order to construct the neural network one needs to rank all the possibilities and only keep the best performing neurons for the next layer. Constructing this ranking homomorphically is not straightforward and not considered at all in our algorithm, hence neural networks will require more complicated algorithms.
When we look purely at the performance of the FV homomorphic encryption scheme, we might consider a residue number system (RNS) variant of the FV scheme as described in [16] to further improve the running time of our implementation. One could also consider single instruction multiple data (SIMD) techniques as suggested in [17] or look further into a dynamic rescaling procedure for FV as mentioned in [6]. These techniques will presumably further decrease the running time of our implementation, which would render our solution even more valuable.
Conclusions
The simple, but effective, iterative method presented in this paper allows one to train a logistic regression model on homomorphically encrypted input data. Our method can be used to outsource the training phase of logistic regression to a cloud service in a privacy preserving manner. We demonstrated the performance of our logistic training algorithm on two real life applications using different numeric data types. In both cases, the accuracy of our method is only slightly worse than standard algorithms to train logistic regression models. Finally, the time complexity of our method grows linearly in the number of covariates and the number of training input data points.
Notes
Funding
This work was supported by the European Commission under the ICT programme with contract H2020ICT20141 644209 HEAT.
Availability of data and materials
The genomic dataset was available upon request during the iDASH competition. Data is still available from the authors upon request and with the permission of the organisers of the iDASH competition of 2017. The financial dataset is not publicly available. They might be made available from the authors upon request and with the permission of the specific undisclosed bank.
About this Supplement
This article has been published as part of BMC Medical Genomics Volume 11 Supplement 4, 2018: Proceedings of the 6th iDASH Privacy and Security Workshop 2017. The full contents of the supplement are available online at https://bmcmedgenomics.biomedcentral.com/articles/supplements/volume11supplement4.
Authors’ contributions
Both authors worked together on the design of the solution. Both authors discussed results and wrote the manuscript together. Both authors have read and approved the manuscript.
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
 1.Naehrig M, Lauter K, Vaikuntanathan V. Can Homomorphic Encryption Be Practical? In: Proceedings of the 3rd ACM Workshop on Cloud Computing Security Workshop, CCSW ’11. New York: ACM: 2011. p. 113–124. http://doi.acm.org/10.1145/2046660.2046682.Google Scholar
 2.Bos JW, Lauter K, Naehrig M. Private predictive analysis on encrypted medical data. J Biomed Inform. 2014; 50:234–243.CrossRefGoogle Scholar
 3.Aono Y, Hayashi T, Trieu Phong L, Wang L. Scalable and Secure Logistic Regression via Homomorphic Encryption. In: Proceedings of the Sixth ACM Conference on Data and Application Security and Privacy, CODASPY ’16. New York: ACM: 2016. p. 142–144. http://doi.acm.org/10.1145/2857705.2857731.Google Scholar
 4.Xie W, Wang Y, Boker SM, Brown DE. PrivLogit: Efficient Privacypreserving Logistic Regression by Tailoring Numerical Optimizers. CoRR. 2016;abs/1611.01170http://arxiv.org/abs/1611.01170.Google Scholar
 5.Kim M, Song Y, Wang S, Xia Y, Jiang X. Secure logistic regression based on homomorphic encryption. IACR Cryptol ePrint Arch. 2018; 2018:14. Accessed 14 Jan 2018.Google Scholar
 6.Fan J, Vercauteren F. Somewhat practical fully homomorphic encryption. IACR Cryptol ePrint Arch. 2012; 2012:144. Accessed 22 Jan 2018.Google Scholar
 7.Lyubashevsky V, Peikert C, Regev O. On Ideal Lattices and Learning with Errors over Rings. J ACM. 2013; 60(6):43–35. https://doi.org/doi:10.1145/2535925.CrossRefGoogle Scholar
 8.Albrecht M. Complexity estimates for solving LWE. 2000. https://bitbucket.org/malb/lweestimator/raw/HEAD/estimator.py. Accessed 15 Aug 2017.
 9.Bonte C, Bootland C, Bos JW, Castryck W, Iliashenko I, Vercauteren F. Faster homomorphic function evaluation using nonintegral base encoding In: Fischer W, Homma N, editors. Cryptographic Hardware and Embedded Systems – CHES 2017. Cham: Springer: 2017. p. 579–600.Google Scholar
 10.Böhning D, Lindsay BG. Monotonicity of quadraticapproximation algorithms. Ann Inst Stat Math. 1988; 40(4):641–663.CrossRefGoogle Scholar
 11.Gershgorin SA. Uber die abgrenzung der eigenwerte einer matrix. Bulletin de l’Académie des Sciences de l’URSS. Classe des sciences mathématiques et na. 1931; 6:749–754.Google Scholar
 12.CryptoExperts. FVNFLlib; 2016. https://github.com/CryptoExperts/FVNFLlib. Accessed 10 May 2017.
 13.Albrecht MR, Player R, Scott S. On the concrete hardness of learning with errors. J Math Cryptol. 2015; 9(3):169–203.CrossRefGoogle Scholar
 14.Böhning D. Multinomial logistic regression algorithm. Ann Inst Stat Math. 1992; 44(1):197–200.CrossRefGoogle Scholar
 15.Bos JW, Castryck W, Iliashenko I, Vercauteren F. Privacyfriendly forecasting for the smart grid using homomorphic encryption and the group method of data handling In: Joye M, Nitaj A, editors. Progress in Cryptology  AFRICACRYPT 2017. Cham: Springer: 2017. p. 184–201.Google Scholar
 16.Bajard JC, Eynard J, Hasan A, Zucca V. A full rns variant of fv like somewhat homomorphic encryption schemes. IACR Cryptol ePrint Arch. 2016; 2017:22. Accessed 22 Jan 2018.Google Scholar
 17.Castryck W, Iliashenko I, Vercauteren F. Homomorphic sim2d operations: Single instruction much more data. IACR Cryptol ePrint Arch. 2017; 2017:22. Accessed 20 Jan 2018.Google Scholar
Copyright information
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.