Portfolio item number 1
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Short description of portfolio item number 2
publications
The Effects of Image Distribution and Task on Adversarial Robustness
Published in CBMM Memo, Under Review at ICML, 2021
This work examines the role of image distribution and task in the adversarial robustness in the resultant model.
Recommended citation: O Kunhardt, A Deza, and T Poggio. The Effects of Image Distribution and Task on Adversarial Robustness, MIT CBMM Memo #116, under review at ICML 2021 https://cbmm.mit.edu/sites/default/files/publications/CBMM_Memo_116_2.pdf
talks
The Effects of Image Distribution and Task on Adversarial Robustness
Published:
Currently, few studies testing the theories of adversarial robustness take into consideration that when doing comparisons across each dataset, both the image distributions (e.g. digits vs objects) and classification task (to classify digits vs objects) are different. To unravel the causal factors of the inherent adversarial robustness of a model, we propose an unbiased metric to compare adversarial robustness and perform a series of experiments. In these experiments, we equalized several training hyperparameters on networks for the MNIST and CIFAR-10 datasets to determine whether the image distribution and task played a role in adversarial robustness. We find that networks trained for digit classification on MNIST digits are more adversarially robust than networks trained to do object classification on CIFAR-10 objects. In addition, to pin-point whether the contribution of adversarial robustness is mainly due to the image distribution or the task, we create a fusion image dataset that overlapped MNIST digits with CIFAR-10 objects such that image statistics were matched and train networks to perform a digit or object classification task. We find that models performing digit recognition on the fusion images were more robust than those performing object recognition, empirically verifying the role of the classification task in the adversarial robustness of a model, independent of the image distribution a network is trained on. Comparing the fusion model performances to their non-fusion counter-parts, we find that image distribution also plays a role.
teaching
Autonomous Navigation Workshop
Workshop, CUNY Hunter College, Computer Science Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 350: Artificial Intelligence
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 127: Introduction to Computer Science
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 350: Machine Learning
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 340: Operating Systems
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 135: Software Analysis & Design 1
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 235: Software Analysis & Design 2
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.
CSCI 335: Software Analysis & Design 3
Undergraduate course, CUNY Hunter College, Computer Science Department, 2021
This is a description of a teaching experience. You can use markdown like any other post.