Homomorphic Model Selection for Data Analysis in an Encrypted Domain^†

Round 1

Reviewer 1 Report

The submitted paper proposes a method of model selection over encrypted data, which presents several implementation algorithms based on bit operations and logarithm functions over encrypted data.

The contribution of this paper would be fair, but the review does not recommend this paper be accepted as it is, because of poor writing and lack of clearness.

First of all, the reviewer found too many grammatical errors. To list only a few of them, i) there are many sentences having an appropriate structure; for example, the second equation in p.5 and Equation (5) are not included in a sentence, and the sentence that includes Equations (4) do not have the period. The word NTRU in p. 2 is not defined, and the phrase ``is denoted by'' in p. 4 should not be in passive voice. Many errors are found in Reference as well. For example, last name of the authors comes first for [7] but first name comes first for [6]. And, first name of the authors should be abbreviated for citation [17] and [20].

Regarding the contribution of this paper and its clearness, the reviewer is confused for a couple of parts. First, there is no proof for the correctness of the proposed algorithm so that it would be hard for readers to follow the technical arguments in this paper. And, there are many parts in which the reviewer was not convinced with the motivation of the proposed schemes. For example, in Section 2.2.1 the reviewer failed to understand why use of floating point consumes more computational complexity than fixed point arithmetic. And, the frequently used phrase ``Our approach is different from the mainstream,'' without explaining why the difference is meaningful, leaves the reader unmotivated, rather skeptical about the contribution.

Author Response

Thank you for your advice and sorry for making you wait long. 

I changed the sentences in my paper through repeated revisions, so I believe that the awkward parts have been reduced from before. 

According to comments, I corrected grammatical errors as well. For example, appropriate structures, NTRU definition, references, etc).

The comment that "There is no proof for the correctness of the proposed algorithm so that it would be hard for readers to follow the technical arguments in this paper." explains in the implementation part. We made sure our inverse matrix algorithm works with the toydataset, and expanded the scope to the Kaggle dataset. Table3 shows how much the difference between the proof value implemented in plaintext and the proof value implemented in ciphertext. 

And the reason "Why use of floating point consumes more computational complexity than fixed point arithmetic" is because of the computational complexity. The more complex the calculation, the slower the operation in the encrypted state, so use fixed-point for efficiency.

This content is introduced in my paper 2.2.1 like this."On the other hand, the major advantage of using a fixed-point representation is that it is a simple integer arithmetic operation with much less logic than floating-point, which improves performance by reducing the bootstrapping procedure on every encrypted bit. Therefore, it is crucial in FHE to reduce, so we only adopt a fixed-point number system in this paper."

Lastly, "Our approach is different from the mainstream" part was to emphasize that our approach is different from that in many papers dealing with homomorphic encryption, but I also deleted this part because I thought it was too much.

There are a lot of shortcomings, but if you have any other corrections, please let me know. Thank you for reading.

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper present homomorphic  model selection methods (Cross-validation and bayesian model selection) in encrypted domain to provide privacy-preserving machine learning. Applying machine learning method to encrypted data is important because in cloud computing services user data typically stays in encrypted. Thus, I think that the methods from this paper have significance in that they are basis for applying machine learning techniques. 

The current problems and the proposed methods are well described. In addition, the results seem to be concrete. Testing on toy dataset and body mass index dataset proves the effectiveness of the proposed methods. Thus, I think that this paper can be published in this journal after a few corrections.

Some typo errors need to be corrected.

ex) P. 2, "and out a new" ->  "and out a new"

P.2, "We have also developed algorithms such as linear regression, logistic regression, KNN, and K-means
through our studies [18,19] based on our HE scheme." -> "Several works have been conducted about linear regression, logistic regression, KNN, and K-means in [18,19] based on the proposed HE scheme."

 

 

Author Response

Thank you for your advice and sorry for making you wait long. 

I changed the sentences in my paper through repeated revisions, so I believe that the awkward parts have been reduced from before. For example, typo errors and sentence structures, etx.

There are a lot of shortcomings, but if you have any other corrections, please let me know. Thank you for reading.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The effort of the authors gone for revising the paper is appreciated, so the reviewer recommends that this paper be accepted.