ICACT20220035 Question.1
Questioner: jypark9740@g.skku.edu 2022-02-15 오전 3:14:48 |
ICACT20220035 Answer.1
Answer by Auhor nishat.mowla@ri.se 2022-02-15 오전 3:14:48
Chrome  Click!! |
| Thank you for your informative papers. I would like to know the detailed environment of hardware and software for experimentation. |
The experiments were conducted on a machine with MacOS Core i9 - 16 GB RAM - 1 TB SSD. The proposed model was created using Python 3.6 in Anaconda environment. Two publicly available datasets were used for evaluation. |
ICACT20220035 Question.2
Questioner: tomayoon@ieee.org 2022-02-15 오후 10:21:56 |
ICACT20220035 Answer.2
Answer by Auhor nishat.mowla@ri.se 2022-02-15 오후 10:21:56
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. |
The research is part of the CyReV project at RISE. The experiments were conducted on a machine with MacOS Core i9 - 16 GB RAM - 1 TB SSD. The proposed model was created using Python 3.6 in Anaconda environment. Some of the relevant libraries used as Numpy, Pandas, scikit learn, etc. Two publicly available datasets were used for evaluation. |
ICACT20220035 Question.3
Questioner: tomayoon@ieee.org 2022-02-15 오후 10:09:41 |
ICACT20220035 Answer.3
Answer by Auhor nishat.mowla@ri.se 2022-02-15 오후 10:09:41
Chrome Click!! |
| Will you tell us what is the major achievement of this paper, and also what is further study issue from this? Thanks in advance!! |
The major achievement of this paper is an explainable model for an automotive intrusion detection system that can be used for the Controller Area Network. To achieve that, we performed extensive experimental analysis on two publicly available CAN bus datasets. We also evaluated the explainability of the proposed solution and further use it to show that reasonable performance can be achieved with a subset of the available features that will also allow reducing the average training time. |
ICACT20220035 Question.4
Questioner: vasaka.vis@mahidol.ac.th 2022-02-15 오후 10:27:23 |
ICACT20220035 Answer.4
Answer by Auhor nishat.mowla@ri.se 2022-02-15 오후 10:27:23
Chrome Click!! |
| Thank you for your presentation. It is an interesting topic.
I have a few questions.
(1) It is my first time to hear about "Explainable model", but when I see your Slide 7 "System Model (2/2)", it is just the normal random forest algorithm.
Which part in this slide that represents the explianable model?
(2) Could you extract the same amount of features from both SAD dataset and ORNL dataset?
If yes, why the average training time of them are different?
Based on your graphs in Figure 3 (Slide 12), for 113 features, the training time of ORNL dataset is almost two times of SAD dataset.
Thank you
|
1) As shown in slide 6 (System model 1/2), the explainable model has two parts: feature exploration and sub-model exploration inspired by the concept of 'exploration as explanation' where explanations are generated from the random forest models used. This generated most contributing features and most contributing sub-models to provide explainability of the performance. 2) Similar feature format was used for both the datasets but the ORNL dataset had a higher number of data points for which it's training time is higher than the SAD dataset. |
ICACT20220035 Question.5
Questioner: nyamai@cc.tuat.ac.jp 2022-02-15 오후 10:18:42 |
ICACT20220035 Answer.5
Answer by Auhor nishat.mowla@ri.se 2022-02-15 오후 10:18:42
Chrome Click!! |
| Thank you for your interesting paper. I have some questions. (1) How would you evaluate whether or not your proposed IDS is explainable? (2) Can the car owner see the output of the IDS and recognize the danger? |
1) The current explainable AI solutions focus on finding feature importance as a measure of explainability. In our proposal, we consider the most contributing features and also the most contributing and least contributing sub-models as a measure of this model's explainability. This explainability of the features and models can be further used to update the parameters of the model. 2) The current solution considers the output to be useful for cyber analysts who have a hard time understanding the decision of the model but as future work, this could be considered as a security app for the car owner that can be run on-board. |
ICACT20220035 Question.6
Questioner: namacabale@gmail.com 2022-02-16 오후 10:47:02 |
ICACT20220035 Answer.6
Answer by Auhor nishat.mowla@ri.se 2022-02-16 오후 10:47:02
Chrome Click!! |
| Dear Author, you have a very good research. Is there an existing technology almost similar to your proposal that has been deployed in real world environment? What does it lack that your work specifically addressed? Thank you and stay safe.
|
Explainable AI is an ongoing research topic and is currently mostly considered for computer vision and similar tasks where explainable AI techniques such as LIME and SHAP are experimented. Explainability is specifically a new dimension but important for ML-based automotive intrusion detection systems. We have compared our results to LIME which also finds the feature importance. In our case, besides finding the feature contributions, we also try to identify sub-model contributions as a measure of explainability. |
ICACT20220035 Question.7
Questioner: yiweimaa@gmail.com 2022-02-16 오후 10:49:40 |
ICACT20220035 Answer.7
Answer by Auhor nishat.mowla@ri.se 2022-02-16 오후 10:49:40
Chrome Click!! |
| Dear Author, you have a very good research. I would like to learn about your explainable AI solution. Whtat's the database you use to be your IDS data? |
We used two datasets: one is the Survival Analysis Dataset for Automobile IDS and the other is the Real ORNL Automotive Dynamometer (ROAD) CAN Intrusion Dataset. Both are publicly available. |
