A Secure Multimodal Biometric Data Protection Framework Using Optimized CNN, GAN-Based Privacy Preservation, and ElGamal Cryptography

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

To ensure the protection of multimodal biometric data transmitted and stored in the IoT and cloud ecosystems, this paper proposes a comprehensive, multi-layered security framework. The authors implement an integrated process comprising three core pillars and demonstrate through experiments that the proposed hybrid process achieves a balance between biometric verification accuracy and data privacy protection.

1. While the authors state in the manuscript that three highly complex components are proposed, the exact data flow and interfaces between the different layers are not clearly shown. This leaves the reader unclear whether the GAN processes the raw image before CNN feature extraction or visualizes it before ELGamal encryption. Could the authors provide a detailed system framework diagram to facilitate reader comprehension?

2. Since ELGamal is a classic public-key cryptosystem based on Diffie-Hellman key exchange, it suffers from ciphertext expansion and a lack of homomorphic properties. Therefore, the authors need to provide a technical justification for choosing ELGamal encryption. For example, although choosing ELGamal encryption incurs additional storage and bandwidth overhead, it is effective for the hybrid process proposed in this manuscript.

3. It is well known that training GANs is very difficult and prone to mode collapse. The authors are advised to provide a specific loss function for training the GAN. 4. The authors clearly state in the manuscript that the optimized GAN achieves high accuracy, but lack a comprehensive comparison with other state-of-the-art (SOTA) networks. It is recommended that the authors add a computational complexity analysis (Big O notation) of this hybrid process to the manuscript.

5. Regarding the manuscript's structure, the authors highlight the optimized CNN in the "Title" and "Abstract," but the specific structural modifications or hyperparameter tuning process that constitutes this optimization is not detailed in Chapter 3. It is recommended that the authors clarify the CNN's layer configuration, filter size, and activation functions.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This article presents a study on the problem of biometric data protection. The topic of this research is highly relevant. The article fills a gap in the existing literature and proposes a secure biometric data protection system (SBDP) that utilizes artificial intelligence and encryption methods. The authors analyzed the problem and provided an overview of previous research. In the study, the authors presented a system that combines an optimized convolutional neural network for extracting distinctive features from multimodal biometric input data, a generative adversarial network, and encryption of biometric features and images using the ElGamal cryptosystem. The study presents experiments on multimodal biometric datasets using TensorFlow, Keras, and PyCryptodome. The results demonstrate the effectiveness of the proposed method compared to CNN, ResNet, Vision Transformer, and ConvGRU models. The SBDP framework proposed by the authors provides a reliable, accurate, and secure solution for real-world applications of biometric authentication and the secure transmission and storage of biometric data.

Figures are used to present the study results. The references cited are relevant to the research topic. The article discusses the limitations of the present study and identifies areas for future research.

The authors should take note of the following comments:

The citations for authors of publications numbered 21–27 need to be verified. Lines 142–199 contain the following references: Vallabhadas et al. [21] proposed…, Sasikala et al. [22] developed…, Sharma et al. [23] proposed…, Srinivas et al. [24] introduced…, Hammad et al. [25] proposed…, Chitrapu et al. [26] proposed…, Aarthi et al. [27] proposed… These references are also listed in Table 1, but we do not see these authors in the reference list under numbers 21–27.

Line 219: It is recommended to add a reference in the text to Table 1.

It is recommended to explain all symbols used in the formulas.

Line 266: It is recommended to explain the symbols H, W, and C in Equation (1).

Line 274: It is recommended to explain the symbols K and N in Equation (2).

Line 295: It is recommended to explain the symbols Φ and L in Equation (5).

Line 315: It is recommended to explain all symbols in Equations (8) and (9).

Line 422: It is recommended to explain h in Equation (20).

Line 503: It is recommended to explain the symbols used in formulas (23)–(29).

Line 600: It is recommended to explain δ for the formula, which is also used in formula (50).

Line 684: It is recommended to explain ε for formula (61).

Line 911: It is recommended to add references to the datasets used.

It is recommended to specify in more detail which ResNet network was used in the study.

It is recommended to include the formulas for the metrics used in the study in the article.

The text of the article provides a detailed comparison of the results of the proposed method with other methods. To better present the comparison of numerical data with the results of studies using other methods, it is recommended to use tables.

It is recommended to list the abbreviations used at the end of the manuscript.

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript proposes a Secure Biometric Data Protection framework that combines optimized CNN-based multimodal biometric feature extraction, GAN-based privacy-preserving synthetic representation generation, ElGamal encryption, and SHA-256-based integrity verification. The topic is relevant to biometric security, privacy preservation, and secure authentication. However, in its current form, the paper has major methodological, theoretical, experimental, and presentation problems.

1. The novelty is not sufficiently established.
    The paper presents the framework as an end-to-end integration of OCNN, GAN, ElGamal encryption, and SHA-256 signature verification. However, most components are standard and are combined in a relatively straightforward pipeline. The claimed contribution therefore appears to be mainly system-level integration rather than a new algorithmic contribution.

2. The “optimized CNN” is not actually defined.
    The manuscript repeatedly uses the term OCNN, but the optimization procedure is unclear. It is not specified whether “optimized” refers to architecture search, hyperparameter tuning, regularization, loss design, feature fusion, or training strategy.

3. The GAN-based privacy claim is not adequately validated.
    The manuscript claims that synthetic biometric representations reduce reconstruction, inversion, and privacy leakage risks. However, the experiments do not include a clear reconstruction attack, inversion attack, membership inference attack, template linkage attack, or adversarial privacy evaluation. Reporting generator and discriminator losses is not sufficient to demonstrate privacy preservation. A privacy-preserving biometric system should be evaluated against explicit attack models.

4. The experimental design is unclear and potentially invalid for multimodal biometrics.
    The paper uses CelebA, a fingerprint dataset, and UBIRIS v2, but it is not clear whether the face, fingerprint, and retina/iris samples correspond to the same individuals. If they are independent datasets with different subject identities, then treating them as a multimodal biometric dataset is problematic.

5. Dataset description contains conceptual inaccuracies.
    UBIRIS v2 is generally an iris dataset, but the paper repeatedly refers to it as retina/retinal images. Iris and retina are different biometric modalities. This is a serious terminology and domain-knowledge issue because retinal scans and iris images have different acquisition processes, anatomical structures, and recognition pipelines.

6. Baseline comparisons are insufficient and possibly unfair.
    The baselines include CNN, ResNet, Vision Transformer, and ConvGRU, but the paper does not specify whether these models use the same preprocessing, same multimodal inputs, same training budget, same hyperparameter tuning, or same feature fusion strategy. Moreover, relevant biometric template protection and privacy-preserving biometric methods should be included as baselines, not only generic deep learning architectures.

7. The framework is not truly end-to-end optimized as claimed.
    The manuscript repeatedly states that feature extraction, GAN transformation, encryption, and verification are jointly optimized. However, encryption and SHA-256 hashing are not differentiable learning components in the usual sense and are not optimized jointly with CNN/GAN parameters. The paper should clarify whether the learning modules and cryptographic modules are trained separately or connected only procedurally.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have carefully reviewed the authors' responses to the reviewers' comments and the corresponding revisions made to the manuscript. I am very satisfied with the revisions provided in response to the points I raised as Reviewer . The authors have addressed all of my concerns thoroughly and thoughtfully. The clarifications added to the text, the additional analysis where requested, and the improved discussion have significantly strengthened the manuscript. The manuscript is now methodologically sound, the results are clearly presented, and the conclusions are well-supported by the evidence. It makes a valuable contribution to the field. Therefore, I recommend acceptance of the manuscript in its current form.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors of the article provided accurate and detailed responses to the comments and made the necessary corrections. The article has been significantly improved. However, the following comments remain:

Line 690: The equation numbering should continue and be 57, not 12.

Line 700: The equation numbering should continue and be 58, not 13.

Line 707: The formula numbering should continue and be 59, not 14.

Line 715: The formula numbering should continue and be 60, not 15.

Line 725: Accordingly, in the following formulas, the numbering should continue and be 61, not 57. Therefore, the equation numbering must also be corrected and continued in the following lines: 735, 741, 748, 757, 800, 805, 819, 812, 818, 822, 834, 837, 839, 841, 1065, 1072, 1077, 1082, 1087, 1091, 1094.

The authors have included detailed numerical results at the end of the article, but references to Appendix A, specifically Tables A1 through A5, need to be added in the text.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

All concerns are addressed.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf