Review Reports - Advancing Federated Learning through Verifiable Computations and Homomorphic Encryption

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Article proposed federated learning and apply homomorphic encryption for federated learning.

(1) Overall, the research is very interesting, worth of consideration for publication.

(2) Federated learning is not clearly presented. Please improve.

(3) homomorphic relationship with Federated learning must be more clearly described.

(4) Experimental section is vague, please improve.

(5) Please share the code so we can investigate more of the algorithms.

Supplement:

The paper proposed federated learning for ZKVM. Overall, the paper is poorly written.

Before the formal review, the paper's English and writing in general must be improved.

For example, the paper did a poor job of describing the motivation for learning. What has been learned?

Why federated learning is used?

Federated Learning is mainly for preserving the privacy of the individual learner. I can not see why it is important here.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors proposed a ZKVM framework/implementation for federated learning.
In short, BGV-FHE is used to provide secure communication between the nodes and aggregator; while the ML computations are verified through the ZK proofs.

The research problem is interesting and the paper is well-written.
However, more discussion is appreciated for the choice of BGV.
For instance, why not choosing BFV or CKKS, as BGV needs to keep track of the noise level at every algorithm step.
Also, it would be nice if the authors can also justify why FHE is selected but not other cryptographic primitives, such as private set intersection protocols:
1) https://doi.org/10.1007/978-3-030-17659-4_5
2) https://link.springer.com/chapter/10.1007/978-3-031-09234-3_29

Besides, the BGV (security) parameter should be given, as it is related to the experiment results in Section 4.

Finally, I hope the authors can show proof of the implementation, e.g., a link to the codes, screenshots, short clips or so on. This is to verify the experiment is indeed done, it does not need to be included in the manuscript.