ICACT20220108 Question.1
Questioner: tomayoon@ieee.org 2022-02-14 오전 10:31:15 |
ICACT20220108 Answer.1
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-14 오전 10:31:15
Chrome  Click!! |
| First of all, thank you for the good research paper and presentation contents. I would like to know more about the research environment used in this study, and also in detail about the equipment & operating systems, development languages and software libraries you used. |
Thank you for your question. The experiments were conducted on a machine with Windows OS, Intel i7-9700 CPU, 32 GB RAM, and NVIDIA GeForce GTX 1650 GPU. We implemented our experiments in MATLAB. |
ICACT20220108 Question.2
Questioner: nay.min@student.mahidol.ac.th 2022-02-15 오후 1:09:54 |
ICACT20220108 Answer.2
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-15 오후 1:09:54
Chrome Click!! |
| Thank you for your presentation. The content is very interesting and I learned something new. As you have received a satisfactory result from your proposed method using the black-box paradigm, do you still plan to test a similar method under a white-box paradigm? |
Thank you for your question. Sorry for the confusion, but our method uses the white-box paradigm.
To answer your question, we're not planning to adjust our method as a black-box paradigm because our method extracts watermarks from the parameters of the host network. |
ICACT20220108 Question.3
Questioner: yiweimaa@gmail.com 2022-02-15 오후 2:57:25 |
ICACT20220108 Answer.3
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-15 오후 2:57:25
Chrome Click!! |
| What are the main factors that cause watermarks of varying degrees to have nearly the same accuracy? |
I'm glad to ask.
It's considered the property of deep learning models. Adding noise in some layers is one of the regularizing methods for improving models. Watermarks function as noise in our method, watermark-embedded models kept their accuracy. |
ICACT20220108 Question.4
Questioner: cho_michael_yoshi@sec.ee.kagu.tus.ac.jp 2022-02-16 오후 12:01:55 |
ICACT20220108 Answer.4
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-16 오후 12:01:55
Chrome Click!! |
| Thank you for your presentation.
The research was very interesting.
Do you have any plans for future research advances, such as increasing the amount of embedded information or other assessing attack resistance? |
Thank you for the question. Yes, we're working on evaluating robustness against fine-tuning and transfer learning. They're often used for deep learning, and affect watermarks by changing parameters in the network.
We also plan to search for more efficient parameters. |
ICACT20220108 Question.5
Questioner: lizhijialzj@gmail.com 2022-02-16 오후 1:00:58 |
ICACT20220108 Answer.5
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-16 오후 1:00:58
Chrome Click!! |
| thanks for your introduction. I would like to know whether to consider other types of neural network compression techniques besides network pruning, and why the accuracy drops sharply when the pruning rate reaches 100%? |
That’s a good point. Actually, I haven’t thought of that so we'll look into it. Thank you for bringing that up.
To answer the second question, it's because of the size of CNN. The smaller the size of the network, the greater the impact of a single layer on its output. Additionally, 100% pruning means that all parameters in the layer turn into zero. Therefore, the accuracy of a small network (like 5-layer CNN) drops significantly if the parameters of one layer are cut by 100%. |
ICACT20220108 Question.6
Questioner: chau@etri.re.kr 2022-02-16 오후 1:13:33 |
ICACT20220108 Answer.6
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-16 오후 1:13:33
Chrome Click!! |
| Dear author, you have a very interesting research. Could we apply your approach for the other domains such as video? Thank you and stay safe. |
Thank you for your question and kindly comment. Yes, as you said, I think it can be expanded to video. This study is based on our previous work about image watermarking, "Collusion-Resistant Watermarking Using Modified Barni Method". Our method just processes matrix, it can apply to other domains as well. |
ICACT20220108 Question.7
Questioner: mryoulin@gmail.com 2022-02-17 오전 12:14:37 |
ICACT20220108 Answer.7
Answer by Auhor kakikura_sayoko@sec.ee.kagu.tus.ac.jp 2022-02-17 오전 12:14:37
Chrome Click!! |
| Dear Presenter, I have learned that your proposed watermark embedding and detecting
algorithm is based on Kang and Iwamura's work, could you please explain how our work had been improved ? |
Thank you for your question. We took over the basic method and adjusted some parameters to fit neural network watermarking. The novelty of this research is that we have introduced our previous method to DNN watermarking. |
