Deepfake Detection Algorithm Based on Dual-Branch Data Augmentation and Modified Attention Mechanism

Round 1

Reviewer 1 Report

The authors have written an article on Deepfake Detection Algorithm Based on Dual-branch Data Augmentation and Modified Attention Mechanism. Please address the following points.

1.     Reference 1 and 3 are the same.

2.     Line no. 37 to 42, the authors have mentioned the problems of existing research. Please justify the statements.

3.     In Line no. 54, you mentioned “weak model robustness”. Please elaborate on the meaning of weakness.

4.     Line no. 95, please expand SVM. Ensure to mention the full form of abbreviations at the first appearance.

5.     In figure 1, “Comprehensive Design Stage” module needs justification. The explanation provided in Lines 150 to 165 is not sufficient. For example, “We introduced the AdamW as the training optimizer in the integrated decision-making stage”. Why did you use AdamW? What is the benefit?

Overall the article is neatly written, but the above points need to be addressed.

Please improve the readability of the article. Avoid using we, I, our model, etc.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper presented two novel methods in improving deepfake images detection including:

AugMix and AugMax data augmentation methods;

The modified attention mechanism module.

The proposed model was tested against commonly used datasets and performs better than the baseline model and other commonly used models.

The paper is well presented and shows the novelty clearly.

The English is clear and understandable. Just a few odd use of words used occasionally, e.g., hardness of training samples. 

Author Response

Point: The English is clear and understandable. Just a few odd use of words used occasionally, e.g., hardness of training samples.

Response: The hardness of training samples refers to the complexity of the training samples. The more difficult the model is to learn, the more complex the samples are and the higher the hardness.

Reviewer 3 Report

I see two basic problems with the papers:

- The first is that there is no real definition of what a fake video means. What changes (for example, histogram equalization, etc.) are accepted, which we do not consider as a fake video? What typical manipulations are involved in making fake videos? How can they be detected? The paper ought to be extended with this information.

- The other important part that can be improved is that not only accuracy is indicated, if we consider binary classifications for the problem, then in this case sensitivity and specificity are also necessary.  Please give these performance metrics about the experiments. If we refer to a multiclass classification problem, then in this case the confusion matrix is also important.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have addressed my comments.

English proofread and polishing is needed. Recheck English langugae.