Privacy-Preserving Federated Learning-Based Intrusion Detection System for IoHT Devices

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

-       This paper proposes a privacy-preserving federated learning farmwork for intrusion detection systems in Internet of Healthcare Things. The proposed framework integrates federated learning with ϵ-differential privacy.  The proposed framework show promising results.

-       For this part in the abstract: “we employ both Deep Neural Network (DNN) and Convolutional Neural Network (CNN) models”, please note that CNN is a type of deep neural networks (DNN). Please consider rewriting these parts here and in other parts in the paper.

-       In section 3.3.1, please provide more info about the used algorithms.

-       Overall, the paper is well written and organized.

Author Response

Comments 1: This paper proposes a privacy-preserving federated learning farmwork for intrusion detection systems in Internet of Healthcare Things. The proposed framework integrates federated learning with ϵ-differential privacy.  The proposed framework show promising results.

Response 1: We thank the reviewer for all the suggestions that improved our work. We have responded to your comments below. Please see our point-by-point responses below.

Comments 2: For this part in the abstract: “we employ both Deep Neural Network (DNN) and Convolutional Neural Network (CNN) models”, please note that CNN is a type of deep neural networks (DNN). Please consider rewriting these parts here and in other parts in the paper.

Response 2: Thank you for pointing out that CNNs are a specific type of DNN. We appreciate your observation and have revised the abstract and relevant parts of the paper to clarify this relationship.

Comments 3:  In section 3.3.1, please provide more info about the used algorithms.

Response 3: We appreciate the comment made by the reviewer. We added more information about the algorithm in Section 3.3.1.

Comments 4: Overall, the paper is well written and organized.

Response: We thank the reviewer for positive feedback.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The paper "Privacy-Preserving Federated Learning-based Intrusion Detection System for IoHT Devices" offers a comprehensive framework combining federated learning with ε-differential privacy to bolster privacy in IDS specifically tailored for Internet of Healthcare Things (IoHT) devices.

1- The integration of ε-differential privacy in federated learning is well-noted; however, the paper should explore more deeply how ε values are chosen and their impact on model performance and privacy trade-offs. Comparative analysis with varying ε values would provide a clearer view of how privacy settings affect the IDS performance.

2-The paper would benefit from a clearer explanation of the training process across different IoHT devices, including how data heterogeneity is handled. Details on the convergence criteria of the federated learning models and the synchronization process across devices should be explicitly stated.

3- More detailed information should be provided about the experimental setup, such as the number of rounds of federated learning executed, the batch sizes used for training, and the specific configurations of the DNN and CNN models.

4- A discussion on the robustness of the proposed models against various types of attacks, such as model poisoning (Like: FED-IIoT paper) or evasion attacks, would be crucial, especially given the sensitive nature of IoHT environments.

5- The security implications of the model aggregation process in federated learning should be analyzed, particularly focusing on whether the aggregation server could become a bottleneck or vulnerability point in the architecture (See and cite this paper: Robust Aggregation Function in Federated Learning).

6- While the paper presents a comprehensive experimental evaluation, it lacks a comparison with state-of-the-art IDS systems that do not use federated learning or differential privacy. Such comparisons could highlight the advantages or potential trade-offs of the proposed approach.

Author Response

Thank you for the review of our manuscript. We have responded to your comments below.

Comments 1: The integration of ε-differential privacy in federated learning is well-noted; however, the paper should explore more deeply how ε values are chosen and their impact on model performance and privacy trade-offs. Comparative analysis with varying ε values would provide a clearer view of how privacy settings affect the IDS performance.

Response 1: Thank you for highlighting this important aspect. In response to your comment, we have conducted additional experiments to explore the effect of varying ε values on the model's performance and the privacy-utility trade-offs. The results of these experiments are now presented in Section 4.5 and 4.6. Information about  the optimal noise value is in the paper in Section 3.3.1.

Comments 2:  The paper would benefit from a clearer explanation of the training process across different IoHT devices, including how data heterogeneity is handled. Details on the convergence criteria of the federated learning models and the synchronization process across devices should be explicitly stated.

Response 2:  Thank you for your insightful comment. We have revised the paper to include a detailed explanation of the training process in our federated learning framework. This includes how we address data heterogeneity across IoHT devices. We have also elaborated on the convergence criteria used to terminate training. These revisions are reflected in Section 3.3.1 of the paper.

Comments 3: More detailed information should be provided about the experimental setup, such as the number of rounds of federated learning executed, the batch sizes used for training, and the specific configurations of the DNN and CNN models.

Response 3:  We appreciate the comment made by the reviewer. We added more information about the algorithm in Section 3.3.1.

Comments 4: A discussion on the robustness of the proposed models against various types of attacks, such as model poisoning (Like: FED-IIoT paper) or evasion attacks, would be crucial, especially given the sensitive nature of IoHT environments.

Response 4:  Thank you for the observation. We improved Section 2. By summarizing the FED-IIoT and comparing it with our work.

Comments 5: The security implications of the model aggregation process in federated learning should be analyzed, particularly focusing on whether the aggregation server could become a bottleneck or vulnerability point in the architecture (See and cite this paper: Robust Aggregation Function in Federated Learning).

Response 5: We are grateful to the reviewer for pointing out this issue. We updated Section 6 by presenting limitations of our work and  detailing more our future work based on the paper:  Robust Aggregation Function in Federated Learning.

Comments 6: While the paper presents a comprehensive experimental evaluation, it lacks a comparison with state-of-the-art IDS systems that do not use federated learning or differential privacy. Such comparisons could highlight the advantages or potential trade-offs of the proposed approach.

Response 6:  Thank you for your valuable suggestion. We would like to point out that our paper already includes a comparison with state-of-the-art centralized deep learning models and the SVM machine learning model for IDS, which do not utilize federated learning or differential privacy. These comparisons are detailed in Section 4.5, Table 3 and Table 5  and highlight key metrics such as accuracy, precision, recall, and F1-score. We believe these comparisons adequately address the reviewer's concern, but we are happy to clarify further or expand on this discussion if needed.

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for the privacy-preserving federated learning framework. Some of my minor suggestions are:

I. Employment of both DNN and CNN models can be explained further.

II. Limitations of the study, if any can be added to enhance proposed framework, SECIoHT-FL.

III. A graph or flowchart depicting the experimental setup can be helpful.

Author Response

Comments 1: Employment of both DNN and CNN models can be explained further.

Response 1:  We appreciate the comment made by the reviewer. We added more information about the algorithm in Section 3.3.1.

Comments 2:  Limitations of the study, if any can be added to enhance proposed framework, SECIoHT-FL.

Response 2:  We appreciate the comment made by the reviewer. We improved the manuscript accordingly to the reviewer's comments and included limitations of our work in Section 6 ( in the future works part).

Comments 3: A graph or flowchart depicting the experimental setup can be helpful.

Response 3:  Thank you for your suggestion. We would like to note that a detailed flowchart depicting the experimental setup is already included in the manuscript (Figure 1). This figure illustrates the steps of data preprocessing, model training with federated learning, the application of differential privacy mechanisms, and evaluation metrics. We believe this flowchart provides a comprehensive overview of the experimental setup and workflow.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors answered my questions.