Robust Zero Watermarking Algorithm for Medical Images Based on Improved NasNet-Mobile and DCT
, Uzair Aslam Bhatti 1, Jing Liu 2, Yen-Wei Chen 3 and Dekai Li 1Round 1
Reviewer 1 Report
The algorithm proposed in this paper provides a robust null watermarking strategy for medical images based on an improved NasNet-Mobile convolutional neural network and discrete cosine transform (DCT) to provide strong security for medical image watermarking. This algorithm uses a deep learning algorithm, a chaotic encryption technique, a perceptual hash algorithm, and a zero watermarking technique to provide security for medical image watermarking information. Experimental results show that the algorithm proposed in this paper can resist a variety of conventional attacks and is excellent against geometric attacks such as rotation, translation, scaling, shearing, etc., and has high robustness. Therefore, the algorithm can also be used for medical images. The algorithm proposed here is very robust compared to other algorithms and uses a deep learning algorithm to extract image features and encrypts the watermark using a more secure method than conventional watermarking methods. Therefore, the algorithm proposed in this paper can be very useful in the field of watermarking medical images.
<Advantages>:
The proposed algorithm exhibits high resistance to various attacks.
The watermark data is encrypted using Chen chaos mapping and Arnold transform shift to enhance the data security.
NasNet-Mobile Convolutional Neural Network is used to train the medical image dataset and extract robust features.
<Disadvantages>:
The paper presents only experimental results and does not provide comparison with other algorithms.
The processing speed of the proposed algorithm may be slow.
<Problems>:
The paper does not provide information about the source and collection method of the data.
The detailed analysis and interpretation of the experimental results is insufficient.
Author Response
Response to Reviewer 1 Comments
Point 1: The paper does not provide information about the source and collection method of the data.
Response 1: We agree with the reviewers. The datasets in this article were obtained from Medical Imaging Park and the American Institutes for Research, and we made additional clarifications, for this, please refer to the highlighted section on page 6 of the article. Please see the attachment.
Point 2: The detailed analysis and interpretation of the experimental results is insufficien.
Response 2: We agree with the reviewers' comments and have added to the analysis of the results in this paper. for this, please refer to the highlighted section on page 11 and 12 of the article. Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The article is interesting and well-prepared. However, there are some remarks that should be implemented in the revised version of the article:
- quality of Fig. 1 should be enhanced,
- the algorithm comparison results visible in Fig. 13 is very valuable. Can you also provide the SOTA comparison for results given in sections 4.2 and 4.3?
- what was the motivation for creating the novel dataset instead of using the existing one (and probably some benchmark to provide a good comparison to SOTA)?
Author Response
Response to Reviewer 2 Comments
Point 1: quality of Fig. 1 should be enhanced.
Response 1: We agree with the reviewer's comments and have revised Fig. 1. for this, please refer to the Fig. 1 on page 4 of the article. Please see the attachment.
Point 2: the algorithm comparison results visible in Fig. 13 is very valuable. Can you also provide the SOTA comparison for results given in sections 4.2 and 4.3?
Response 2: Thank you for your comments. The algorithm comparison results in Figure 13 are based on the more representative experimental results in Sections 4.2 and 4.3 of this paper compared with other watermarking algorithms, which can present readers with clearer results.
Point 3: what was the motivation for creating the novel dataset instead of using the existing one (and probably some benchmark to provide a good comparison to SOTA)?
Response 3: Thank you for your question. In order to make the NasNet-Mobile-DCT algorithm proposed in this paper applicable to medical images of different parts of the human body to resist various types of attacks, so instead of using the existing datasets, we selected 500 medical images from the Medical Image Park and the American Institutes for Research website and performed data augmentation to create datasets for the categories of brain, abdomen, thorax, bones, and muscles. The experimental results show that We achieved good experimental results using a dataset we created ourselves.
Author Response File: Author Response.pdf
