Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. This is a combination of existing ideas without clear conceptual novelty. You MUST reframe your novelty more sharply.
2. Right now, your method is engineering-driven, not theory-driven. Pls add design proof subsection. For example Pixel-level with noise artifacts; local with blending boundaries and so on.
3. From page 6–7 MDFP Block  looks like Module Aggregation. This is just multi-branch attention fusion. Pls add dynamic weighting explanation and add cross-branch interaction.
4. CGDC block is strong — but UNDER-SOLD. Currently it sounds like feature fusion, pls reframe it as core novel contribution.
5. Missing comparison with SOTA transformers. Pls add models belongs to vision Transformer-based models.
6. Missing robustness tests. Pls add compression (JPEG); noise; blur and low resolution

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript addresses the important and timely problem. The application-oriented motivation is a strength of the paper. The proposed framework appears technically coherent, and the emphasis on efficiency rather than only benchmark accuracy is valuable.

My remarks:

• The title is relevant and technically aligned with the manuscript, but it is somewhat generic.
• The abstract is too general and does not include quantitative results.
• The keywords could be improved. Some are too broad, and the set does not fully capture the forensic and deployment-oriented aspects of the paper.
• The experimental setup is described too generally. Reporting only the use of an NVIDIA RTX 4090 GPU and PyTorch is insufficient. The authors should specify the software and hardware environment in more detail, including Python, PyTorch, and CUDA versions, as well as key libraries used in the implementation.
• Since the paper strongly emphasizes efficiency, the computational evaluation should go beyond reporting only the number of parameters and FLOPs. At minimum, inference time per frame/image and throughput should also be provided under clearly defined conditions.
• The architectural and hyperparameter choices appear arbitrary. The authors do not sufficiently explain why specific parameters or design settings were selected. It should be clarified whether these choices were based on preliminary experiments, literature evidence, or implementation constraints.
• The readability of Figures 2, 3, and 4 should be improved (larger fonts, etc.).
• Figure 5 should be enlarged. The journal layout allows better use of horizontal space.
• As the experiments involve multiple datasets, the qualitative results should include more examples from different datasets. This would better demonstrate robustness/performance across different forgery scenarios.
• The authors should also include examples of incorrect detections, including false positives and false negatives.
• The Data Availability Statement addresses the datasets, but there is no information about code availability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

No more comments need

Author Response

We sincerely appreciate the time and effort that Reviewer 1 has dedicated to improving our manuscript. We are grateful for the constructive comments and valuable suggestions, which have significantly helped us enhance the quality of this work.   Thank you very much for your recognition and approval.

Reviewer 2 Report

Comments and Suggestions for Authors

The revised manuscript has improved substantially, and the authors have addressed most of my previous comments.

That said, a few concerns still remain:

1. The newly introduced theoretical framing, including terms such as “physical prior-driven”, “authentication gap”, and especially “design proof”, appears somewhat overstated. In its current form, this part reads more like a post hoc conceptual justification of architecture than a demonstrated theoretical foundation. I would recommend softening this terminology and presenting it more explicitly as design motivation or architectural rationale rather than as proof.
2. The response letter contains a noticeable editorial inconsistency - although the manuscript is submitted to Electronics, the authors state that they hope the revised version will be suitable for publication in Computers, Materials & Continua. This suggests that the response may have been prepared using another template or a previous rebuttal file without sufficient revision. While this does not directly affect the scientific content, it gives an impression of insufficient editorial care.
3. Some parts of the revised manuscript still rely on references and comparisons that do not fully reflect the most recent developments in deepfake detection. Given the fast pace of this field, I believe the paper would benefit from at least a brief discussion of the latest trends and recent directions in the literature (e.g. “deepfake detection” search in Google Scholar with a time limit from 2025).
4. Finally, some parts of the revised text still use language that is slightly too strong or promotional, especially in the interpretation of Grad-CAM visualizations and in the description of the method’s conceptual novelty. A more neutral scientific tone is advised.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf