Do you have any detailed documentation?

[TimefliesG]

Looking forward to your reply. Thank you.

[TimefliesG]

Hello, author! Thank you for your work. The problem I encountered is: Many .pt model files are obtained through public file training, but the next step is not given in README.md. Can we get the blurry picture as in the paper through the trained model? Or can the detailed file describe the next use of the model?And in the public file, gradcam.py is missing test_all_experiment_models,combine_models_for_analysis. Will the absence of these two files affect the results of the experimental test?

[tremblerz]

Hi @TimefliesG !
Thanks for identifying the issues. I would recommend using the src folder instead for generating the pt files. For obtaining the reconstructions, you can use the decoder_attack module and get those blurry (or non-blurry depending upon the method you use) images.

There are some images that might not be possible to get through the src directory but you can use the following reconstruction attack for that purpose.

[tremblerz]

If you can tell your exact requirement, I might be able to help more.

[TimefliesG]

Thank you for your answer. @tremblerz
I will use src to generate .pt. If I encounter problems again, I will actively ask you for advice,
Best wish!

[TimefliesG]

Are the defense mechanisms you mentioned in your DISCO paper not included in this benchmark? If I want to do research in this area, can you open source related content?

[tremblerz]

Hi @TimefliesG can you mention which specific ones are you referring to? We have most of them integrated and a very few left to be merged from the PRs

[TimefliesG]

Hello @tremblerz Sorry for not replying to you in time due to time difference and classes etc. I am referring to the algorithm part (algos folder), such as disco.py in the algorithm, this file is empty, is it the dynamic and invariant sensitive channel fuzzing method for deep neural networks mentioned in the DISCO paper?

[tremblerz]

We should have disco merged very soon. If you wanna use it right away you can find it here https://github.com/splitlearning/InferenceBenchmark/blob/main/public/config.json, run python main.py in that directory and modify the config accordingly

[TimefliesG]

Thank you for your work on this, I'll keep following and learning from this work ！

[TimefliesG]

Hi @tremblerz In the data folder, dataset_utils.py does not contain UTKFace, but it is referenced in line 72 of loaders.py. Was this part not added?

[tremblerz]

I think it might have been kicked out due to some pull request, we will try to bring it back

[TimefliesG]

In public, using gram-cam.py, the necessary files are missing: test_all_experiment_models and combine_models_for_analysis. At line 510, output_dir is missing. Could you mind provide the corresponding code for learning, thanks!