Weekly Updates for Krishna Khadka

Hi Group,

Milestone: I would like to set up small milestones and major milestone for the winter break. I will have them sent over my channel by Thursday evening,

Last week, I haven’t been able to make much progress due to not feeling well and had some gradings to do.

My goal for this week is to have a timeline and get some experimental results.

you brought up the issue of research productivity last time. i would suggest you look back and also forward to see if there is anything you can do to address it.

Hi Group,

Last week I was going through some papers to get better understanding of the implementation of adversarial samples for tabular data.

The papers include:

• Decision Boundaries of Deep Neural Networks – understanding how adversarial training mitigate the adversarial attacks.

• HopSkipJumpAttack: A Query-Efficient Decision Based Attack – a black box based attack.

• Constrained Adaptive Attack: Effective Adversarial Attack Against Deep Neural Networks for Tabular Data – relatively latest approach to generating adversarial samples for tabular data.

I was planning to investigate the implementation of codes for each of these papers, however, I wasn’t able due to feeling unwell.

I do realize the need to stay on track towards the project. I plan to have daily and weekly targets for myself. Since, I have not completely recovered, I would like to join today’s meeting by teams. Thank you.

sounds good. hope you will recover soon.

Hi All,

My current action items include:

1. Generate adversarial samples for tabular data. I’m exploring existing techniques to implement it. The existing approach I used for adversarial image generation doesn’t directly apply to tabular data. I was also thinking if I should do some experiments on simpler dataset for images such as MNIST, FashionMNIST.

1. I also spent some time exploring the use of combinatorial testing to improve bayesian optimization. I’m trying to set up experiments where I use t-way set of hyperparameters for initial exploration and then apply exploitation on promising regions identified.

in general, when you explore a new idea, try to focus on the core idea. that is, you need to simplify other less essential aspects. in your case, it seems much easier to explore the new idea with tabular data than image data.

also as you noted, i do notice a productivity decline recently. i would encourage you take a moment and reflect what is causing this and what you can do to change the momentum.

Hi Group,

Action items completed last week:

1. Created a blog post website for my research project.

1. https://krishnakhadka200416.github.io/

1. After applying the adversarial approach and identifying key segments in images, I am encountering an issue where regions unimportant are being marked as important.. I’m having difficulty to get some good results. I need to do a thorough implementation check to make sure everything is correct. For example, the image below shows 5 important super-pixels identified by the approach, where the model identifies regions other than dog as important.

I would like to request a leave from tomorrow’s meeting due to attending a relative’s wedding reception in Austin.

the blog webpage looks good. yes, try to verify the validity of each step. also try tabular data.

Hello All,

Milestone: Select a new milestone date for paper submission. Set up internal deadlines to stay on track.

Currently, I’m working on integrating our approach, the adversarial generation, to LIME’s code framework for tabular and image dataset and then do the evaluation with LIME. I will present the evaluation results on Friday.

For the survey paper, I’ve completed an outline and looking for papers within the outline.

Hello Group,

Milestone: Submit a paper on Nov 10

Action items performed last week:

1. After Friday’s discussion, I have made some changes to the code to construct a logistic regression model for explaining image prediction.

1. Working on outlining for survey paper, identifying related papers.

can you outline the approach step by step, and then try to stick to the approach to see how each step works? also tabular data may be easier to prototype in the beginning.

Milestone: Submit a paper on Nov 10

The action items performed this week:

1. I implemented the approach made last week.

1. Generated adversarial of adversarial.

1. Trained a local model on those adversarial.

1. Generated a super pixel based feature importance.

1. I’m now working to setup evaluation with other approaches.

1. I have also been working on outlining for the survey paper on surrogate model construction. I will try to have a basic outline by Friday.

sure @Sunny

Hi Group,

Milestone: Make a paper submission Nov 9

Action items completed this week.

1. I’m exploring how to sample near the closest decision boundary, guided by normal to the decision boundary in the latent space identified, for constructing local surrogate model. Currently, I’m understanding how to construct a decision boundary in the latent space. Then sample along the normal to the decision boundary from the instance to explain. This is to ensure the samples used to train local model are realistic and more informative.

1. I’m also exploring more on constructing ensemble of decision trees using t features for reducing overfitting and better generalization during inference.

try to come up with a baseline approach for how to construct a local surrogate model. see my comments to pujan’s update

Hi All,

Milestone: Trained a local surrogate model on images generated using diffusion model Oct 8 :white_check_mark:

Major Milestone: Submit a paper to CAIN 2025 – Nov 9

This week I have been focusing on different approaches sample in the latent space. I’m exploring on how I can apply or improve one of these approaches for improving my project.

what are you findings about the local surrogate model you trained?

I haven’t yet done the evaluation part. However, I generated 100 perturbations using the diffusion model and trained a local CNN model and explain using integrated gradients, similar to LRP, approach.

Hi Group,

Milestone: Develop a local model using images generated by the diffusion model. – Oct 8

Following our discussion last week, I have been reviewing literature on approaches to improve the samples generated by the diffusion model.

My focus is primarily on research that utilizes the latent space of the diffusion model to create perturbations, as I believe this approach will be relevant for our project. I am also in process of implementing their code to generate images.

The paper is: https://ieeexplore.ieee.org/document/10581912

the goal of this project is to improve the quality of the instances sampled for building high-fidelity local models. the quality will be in terms of realism of individual instances and representativeness of all the instances.

Two specific things we discussed but I have not got a clear idea is: (1) what are the existing sampling approaches for building local models? are there fundamental differences between those for high accuracy and those for high fidelity? (2) are there any existing work on sampling from latent space to build local surrogate models? Can you make a presentation on this today or Friday’s meeting? Please try to focus on these two things.

I will give a presentation on Friday

Hi All,

Milestone: Build a local surrogate model using images generated using diffusion model – 9/27

Action Items Performed Last Week:

• I applied perturbations to t superpixels within images. This approach allows us to pinpoint superpixels critical to the model’s predictions by observing the decrease in prediction probability when these superpixels are altered, while maintaining visual similarity to the original image and unlike using black masking.

1. The latent stable diffusion model applies a diffusion process within an autoencoder framework. Given the large size of the original images, they are encoded into a latent space where the diffusion process is applied. The decoder then reconstructs the image back to its original form. This approach generated visually similar images that had same and different labels compared to the original labels.

• Class ID 264: 68 instances

• Class ID 251: 9 instances

• Class ID 162: 1 instance

• Class ID 222: 13 instances

• Class ID 207: 5 instances

• Class ID 208: 3 instances

• Class ID 167: 1 instance

i don’t quite understand item 1. item 2 seems to be exactly what we would like. please explain the results in today’s meeting.

Hello All,

1. I have been looking into sampling ways for building local surrogate models in existing literature.

1. After Fridays, discussion, I tried using stable diffusion model to generate some noise in the input image. The process is time consuming.

1. I have also been working on building a local surrogate model using t-way sampling of the IPM generated from the metadata.

I plan on continuing my above tasks.

can you put a target date on the milestone? also did you finish your survey/presentation on local surrogate model?

Hi All,

Milestone: Prepare a presentation on local model building for friday’s meeting.

I have been collecting and reviewing research papers regarding local surrogate model building.

I have prepared a presentation for tomorrow where I will present an approach to build local surrogate model using t-way tests and interpret a black box prediction for tabular data.

look forward to the presentation tomorrow