Adversarial ML · NYU Tandon · 2025
Adversarial Robustness in Deep Neural Networks
An empirical study of adversarial training, validating Madry et al. (2018): train MNIST classifiers against a Projected Gradient Descent (PGD) adversary and measure how robustness generalizes across attack families.
Approach
The PGD-trained model held ~90% accuracy under PGD and FGSM attacks — but the more interesting result is where it failed: Carlini & Wagner (C&W) optimization-based attacks and distribution shifts both punched through, a concrete demonstration that robustness against one attack family is not robustness in general.
A significant chunk of the work was engineering: re-engineering TensorFlow 1.x research code to 2.x so CleverHans and the Adversarial Robustness Toolbox (ART) could run in one reproducible pipeline.
Highlights
- ~90% adversarial accuracy under PGD/FGSM after adversarial training
- Demonstrated C&W and distribution-shift weaknesses in PGD-trained models
- TensorFlow 1.x → 2.x migration of the full experimental stack
- Reproducible attack evaluation across CleverHans and ART