Lightweight Photo-Response Non-Uniformity Fingerprint Extraction Algorithm Based on an Invertible Denoising Network
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper "Lightweight PRNU fingerprint extraction algorithm based on 2 an invertible denoising network" proposes an algorithm (Invertible Denoising Network (InvDN) to enhance the quality of PRNU fingerprint extraction and reduces the number of model parameters in order to be deployed on low processing power systems.
The Abstract and keywords are relevant.
Paper goal and structure are outlined in the introduction section. It also contains the brief literature review.
The algorithm and the two components are presented in Chapter 2.
The experimental setup and the used datasets are presented in Chapter 3. The results are compared with other solutions and an advantage is observed, especially for lower resolution images. The classifier accuracy is assessed in Chapter 3.2.
Finally, conclusions are presented outlining algorithm advantages and future development.
The use of English Language is good.
Observations:
Please check reference style to be consistent (reference 12)
References are novel, possibly more should be investigated, especially designed algorithm results in case of counter-forensic transformations, as presented in:
Martín-Rodríguez, F.; Isasi-de-Vicente, F.; Fernández-Barciela, M. A Stress Test for Robustness of Photo Response Nonuniformity (Camera Sensor Fingerprint) Identification on Smartphones. Sensors 2023, 23, 3462.
Authors state: "Furthermore, the algorithm employed in this study utilizes a significantly lower parameter quantity compared to other algorithms of similar nature, resulting in efficient PRNU fingerprint extraction performance while conserving memory usage."
Is there a memory comparison available with the other algorithms (advantage for using fewer parameters)?
What about the processing time, are these comparable with other algorithms?
Author Response
Comments 1: Please check reference style to be consistent (reference 12)
Response1:Thank you very much for taking the time to review this manuscript. Following your suggestion, we check the reference 12.
Comments 2: References are novel, possibly more should be investigated, especially designed algorithm results in case of counter-forensic transformations, as presented in:
Martín-Rodríguez, F.; Isasi-de-Vicente, F.; Fernández-Barciela, M. A Stress Test for Robustness of Photo Response Nonuniformity (Camera Sensor Fingerprint) Identification on Smartphones. Sensors 2023, 23, 3462.
Response 2: We sincerely appreciate the reviewer’s valuable suggestion to investigate more references,We have added a reference section at the end of the article.
Comments 3:
"Furthermore, the algorithm employed in this study utilizes a significantly lower parameter quantity compared to other algorithms of similar nature, resulting in efficient PRNU fingerprint extraction performance while conserving memory usage."
Is there a memory comparison available with the other algorithms (advantage for using fewer parameters)?
What about the processing time, are these comparable with other algorithms?
Response 3 : Thank you very much for your advice. I apologize for not being able to conduct experiments to reduce the number of model parameters (memory consumption) due to the current experimental conditions, but we have theoretically analyzed the network structure for this purpose.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsSummary of the Work
A lightweight photo-response non-uniformity (PRNU) fingerprint is an optimized and efficient version of the traditional PRNU noise pattern, tailored for applications where computational resources or time are limited. The authors introduce a novel PRNU fingerprint extraction algorithm leveraging an invertible denoising (InvDN) network. This approach enhances the accuracy of PRNU fingerprint extraction while providing a more efficient solution, making it particularly suitable for real-world forensic applications.
Key Findings
i) Experimental results confirm that InvDN-based methods outperform other algorithms in PRNU fingerprint extraction accuracy.
ii) The proposed algorithm is more easily deployable on small devices than state-of-the-art PRNU fingerprint extraction methods.
General Remarks
- The list of references should be completed, possibly with the help of the suggestions given below.
- As the authors also recognize, while the algorithm proposed in this study offers a structured and lightweight method for PRNU fingerprint extraction and similarity assessment, it also has some inherent limitations that need to be addressed for practical implementation (see the suggestions below).
- A lightweight PRNU fingerprint is an optimized, efficient version of the traditional PRNU noise pattern designed for applications where computational resources or time constraints are critical. However, lightweight approaches may be less robust against anti-forensic techniques or heavy post-processing (e.g., compression, cropping). The limitations of this technique should clearly be highlighted.
- While invertible denoising network-based PRNU fingerprint extraction algorithms show significant promise and could outperform state-of-the-art methods in many scenarios, their superior performance cannot be universally guaranteed (see the suggestions below).
- The limitations mentioned by the authors ("slow speed and low efficiency in actual operation in the experimental environment") are not the only potential limitations of the PRNU fingerprint extraction algorithm described in the flowchart. While the authors emphasize deployment challenges and operational inefficiencies, additional limitations can be inferred based on the algorithm's structure and potential real-world conditions (see the suggestions below).
The following suggestions aim to fill some gaps.
Suggestions
1) It should be clear from the outset that simplifying the network to make it lightweight may inadvertently compromise its ability to model complex PRNU patterns, especially in compressed images. In my opinion, this should be highlighted.
2) PRNU fingerprints can be deliberately tampered with using anti-forensic attacks or image manipulation techniques. It remains unclear how well invertible denoising networks resist such attacks compared to conventional methods. Furthermore, while lightweight networks are computationally efficient during inference, the training phase might still be computationally expensive and resource-intensive. The authors are asked to discuss the above issues.
3) The variability of PRNU patterns across different devices and sensor types may pose a challenge. Although deep learning models can adapt, they may fail to generalize to entirely unseen devices or camera models without retraining or fine-tuning. Additionally, the performance of invertible networks can be sensitive to the choice of architecture, loss functions, and hyperparameters, which may require significant experimentation to optimize.
4) Fig.1 reports the algorithm framework proposed in this study. We may object that the performance of invertible networks can be sensitive to the choice of architecture, loss functions, and hyperparameters, which may require significant experimentation to optimize. Moreover, simplifying the network to make it lightweight may inadvertently compromise its ability to model complex PRNU patterns, especially in noisy or compressed images. The authors are asked to reply to the above concerns.
5) The authors propose an algorithm that offers a structured and lightweight method for PRNU fingerprint extraction and similarity assessment. However, we may object that the effectiveness of the initial filtering directly impacts the quality of the extracted PRNU fingerprint. If the filtering is suboptimal, critical PRNU features may be lost or distorted, reducing accuracy. Furthermore, enhancing the fingerprint can amplify unwanted noise or artifacts, potentially leading to false positives or decreased specificity in similarity calculations. The authors are asked to dispel these possible objections.
6) As mentioned in the above section, the robustness to real-world variability requires a deeper discussion. More specifically, compression, resizing, cropping, and other common image processing operations may weaken the extracted PRNU fingerprint, reducing the reliability of similarity assessments. Besides, PRNU patterns vary significantly between devices and can overlap in similar sensors or camera models, potentially causing misclassification. Finally, the
environmental factors may negatively affect the efficiency of the authors' algorithm. Indeed, lighting conditions, motion blur, or sensor noise during image capture may further degrade the PRNU pattern. The above-mentioned issues should be discussed very carefully.
7) To summarize, the authors propose a lightweight algorithm for extracting PRNU fingerprints, using an invertible denoising network (InvDN) for source camera identification. In this context, the following questions immediately arise: “Can advanced adversarial attacks or anti-forensic techniques intentionally alter PRNU fingerprints, undermining the reliability of similarity calculations?' Additionally, “Is the proposed algorithm vulnerable to injected noise or pattern suppression methods aimed at obscuring the PRNU fingerprint?”
8) The authors claimed that the major limitations of their PRNU fingerprint extraction algorithm are the slow speed and low efficiency in actual operation in the experimental environment. However, the overall limitations of the algorithm are broader, encompassing technical, computational, and robustness challenges. For instance, ineffective filtering may lead to incomplete separation of the content and noise, compromising the extracted PRNU fingerprint. Furthermore, combining multiple processing stages can lead to increased computational overhead, particularly for high-resolution images, which may reduce the algorithm's efficiency. Moreover, excessive reliance on enhancement techniques could heighten sensitivity, potentially amplifying artifacts or distortions in the PRNU fingerprint and compromising its reliability. The authors are asked to mention the above limitations.
Conclusions
This study is intriguing and timely. However, certain aspects require further clarification and a more detailed discussion. The authors are encouraged to address the above suggestions and expand the cited literature as appropriate.
Author Response
Comments 1: It should be clear from the outset that simplifying the network to make it lightweight may inadvertently compromise its ability to model complex PRNU patterns, especially in compressed images. In my opinion, this should be highlighted.
Response 1:I agree, and in the revised conclusion I have addressed the potential limitations of this model in compressing images.
Comments 2: PRNU fingerprints can be deliberately tampered with using anti-forensic attacks or image manipulation techniques. It remains unclear how well invertible denoising networks resist such attacks compared to conventional methods. Furthermore, while lightweight networks are computationally efficient during inference, the training phase might still be computationally expensive and resource-intensive. The authors are asked to discuss the above issues.
Response 2:Thank you very much for your suggestion, and I have included a discussion about training resources related to anti-forensics in the article.
Comments 3:The variability of PRNU patterns across different devices and sensor types may pose a challenge. Although deep learning models can adapt, they may fail to generalize to entirely unseen devices or camera models without retraining or fine-tuning. Additionally, the performance of invertible networks can be sensitive to the choice of architecture, loss functions, and hyperparameters, which may require significant experimentation to optimize.
Response 3:Thank you very much for your suggestions. I am sorry that I cannot supplement the training experiments about network structure, parameters, loss functions, etc. due to the limited experimental conditions (including equipment, datasets, etc.).
Comments 4:Fig.1 reports the algorithm framework proposed in this study. We may object that the performance of invertible networks can be sensitive to the choice of architecture, loss functions, and hyperparameters, which may require significant experimentation to optimize. Moreover, simplifying the network to make it lightweight may inadvertently compromise its ability to model complex PRNU patterns, especially in noisy or compressed images. The authors are asked to reply to the above concerns.
Response 4:Thank you very much for your suggestions. The limitations of this study in terms of model training have been explained in reply 3. In addition, the lightweight reversible denoising network adopted in this algorithm achieves the effect of reducing model parameters due to the characteristics of the reversible network itself, and the extraction quality of PRNU fingerprint mainly depends on the degree of noise removal in the filtering stage, so I believe that the lightweight feature of this network will not cause additional loss of PRNU fingerprint.
Comments 5:The authors propose an algorithm that offers a structured and lightweight method for PRNU fingerprint extraction and similarity assessment. However, we may object that the effectiveness of the initial filtering directly impacts the quality of the extracted PRNU fingerprint. If the filtering is suboptimal, critical PRNU features may be lost or distorted, reducing accuracy. Furthermore, enhancing the fingerprint can amplify unwanted noise or artifacts, potentially leading to false positives or decreased specificity in similarity calculations. The authors are asked to dispel these possible objections.
Response 5:Thank you very much for your suggestion. The enhancement module of the algorithm in this paper may indeed expand the irrelevant noise remaining in the noise residual after filtering. The research focus of this paper is mainly to extract PRNU fingerprints with higher purity as much as possible by improving the performance of the filter in the filtering stage.
Comments 6: As mentioned in the above section, the robustness to real-world variability requires a deeper discussion. More specifically, compression, resizing, cropping, and other common image processing operations may weaken the extracted PRNU fingerprint, reducing the reliability of similarity assessments. Besides, PRNU patterns vary significantly between devices and can overlap in similar sensors or camera models, potentially causing misclassification. Finally, the environmental factors may negatively affect the efficiency of the authors' algorithm. Indeed, lighting conditions, motion blur, or sensor noise during image capture may further degrade the PRNU pattern. The above-mentioned issues should be discussed very carefully.
Response 6:We greatly appreciate the reviewer's suggestion to further supplement the experiments simulating real-world scenarios.Thank you for your suggestion. According to reference 1-8, PRNU fingerprint is caused by the defect of silicon element in the fabrication process of imaging device sensor, so it is unique for different imaging devices. In addition, the experiment of this paper has carried out 128*128,256*256 and 512*512 cropping processing on the images in the experimental data set, and the results prove that the performance of the algorithm in this paper can remain relatively stable under image cropping. As for the possible performance degradation caused by image compression, I have discussed in the conclusion of the revised paper. In fact, the Daxing and Dresden datasets used contain images from a variety of real-world scenarios, already accounting for variations in lighting and image texture. We apologize for not clarifying this point earlier. We provide a supplementary explanation in Section 3.1.
Comments 7:To summarize, the authors propose a lightweight algorithm for extracting PRNU fingerprints, using an invertible denoising network (InvDN) for source camera identification. In this context, the following questions immediately arise: “Can advanced adversarial attacks or anti-forensic techniques intentionally alter PRNU fingerprints, undermining the reliability of similarity calculations?' Additionally, “Is the proposed algorithm vulnerable to injected noise or pattern suppression methods aimed at obscuring the PRNU fingerprint?”
Response 7:Thank you for your suggestions. The discussion of anti-forensic techniques and whether noise suppression and noise injection processing will affect the performance of the algorithm in this paper is given in the conclusion of the paper.
Comments 8: The authors claimed that the major limitations of their PRNU fingerprint extraction algorithm are the slow speed and low efficiency in actual operation in the experimental environment. However, the overall limitations of the algorithm are broader, encompassing technical, computational, and robustness challenges. For instance, ineffective filtering may lead to incomplete separation of the content and noise, compromising the extracted PRNU fingerprint. Furthermore, combining multiple processing stages can lead to increased computational overhead, particularly for high-resolution images, which may reduce the algorithm's efficiency. Moreover, excessive reliance on enhancement techniques could heighten sensitivity, potentially amplifying artifacts or distortions in the PRNU fingerprint and compromising its reliability. The authors are asked to mention the above limitations.
Response 8:Thank you for your advice, and I will elaborate on these limitations in the future outlook (conclusion) section of the paper.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have addressed the suggestions outlined in my previous report. However, despite their claims in the 'Author's Reply,' the revised manuscript does not introduce any significant new contributions. Nonetheless, I find the manuscript interesting and timely, and I therefore recommend it for publication.