Nabeel Hingun

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2023-153

May 12, 2023

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-153.pdf

Part models have been shown to be an effective way to increase the robustness of deep learning models to adversarial examples. While part models have successfully been applied to small datasets like PartImageNet, obtaining the necessary segmentation labels can be expensive and time-consuming. In order to scale part models to larger datasets, it is crucial to find ways to obtain cheaper labels. In this work, we explore some of the challenges that may arise when scaling up part models. First, we investigate ways to reduce labeling costs by using part bounding box labels instead of segmentation masks, while still providing additional supervision to models. Second, we evaluate the performance of part-based models on a more diverse and larger dataset. Our work provides valuable insights into the key challenges that need to be addressed in order to scale up part models successfully and achieve adversarial robustness on a larger scale. The code is publicly available at https://github.com/nab- 126/adv-part-based-models.

Advisors: David A. Wagner

---

BibTeX citation:

@mastersthesis{Hingun:EECS-2023-153,
    Author= {Hingun, Nabeel},
    Editor= {Wagner, David A.},
    Title= {Scaling Part Models: Challenges and Solutions for Robustness on Large Datasets},
    School= {EECS Department, University of California, Berkeley},
    Year= {2023},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-153.html},
    Number= {UCB/EECS-2023-153},
    Abstract= {Part models have been shown to be an effective way to increase the robustness of deep learning models to adversarial examples. While part models have successfully been applied to small datasets like PartImageNet, obtaining the necessary segmentation labels can be expensive and time-consuming. In order to scale part models to larger datasets, it is crucial to find ways to obtain cheaper labels. In this work, we explore some of the challenges that may arise when scaling up part models. First, we investigate ways to reduce labeling costs by using part bounding box labels instead of segmentation masks, while still providing additional supervision to models. Second, we evaluate the performance of part-based models on a more diverse and larger dataset. Our work provides valuable insights into the key challenges that need to be addressed in order to scale up part models successfully and achieve adversarial robustness on a larger scale. The code is publicly available at https://github.com/nab- 126/adv-part-based-models.},
}

EndNote citation:

%0 Thesis
%A Hingun, Nabeel 
%E Wagner, David A. 
%T Scaling Part Models: Challenges and Solutions for Robustness on Large Datasets
%I EECS Department, University of California, Berkeley
%D 2023
%8 May 12
%@ UCB/EECS-2023-153
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-153.html
%F Hingun:EECS-2023-153