https://github.com/xnul/pgd-reevaluation

https://github.com/xnul/pgd-reevaluation

Last synced: 7 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/xnul/pgd-reevaluation
• Owner: xNul
• Created: 2021-04-27T01:01:40.000Z (over 4 years ago)
• Default Branch: main
• Last Pushed: 2021-04-27T01:05:41.000Z (over 4 years ago)
• Last Synced: 2025-01-21T17:26:54.098Z (9 months ago)
• Language: Python
• Size: 27.3 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# "Towards Deep Learning Models Resistant to Adversarial Attacks" Re-Evaluation

Code used to re-evaluate the Projected Gradient Descent (PGD) statistics given in "Towards Deep Learning Models Resistant to Adversarial Attacks".

Credit to the original paper.

# Results

MNIST Attack PGD k=40 randomrestart=1

natural: 99.17%

adversarial: 0.00%

avg nat loss: 0.0315

avg adv loss: 46.6018

MNIST Defense PGD k=40 randomrestart=1

natural: 98.53%

adversarial: 94.23%

avg nat loss: 0.0409

avg adv loss: 0.1763

MNIST Defense PGD k=100 randomrestart=1

natural: 98.53%

adversarial: 92.70%

avg nat loss: 0.0409

avg adv loss: 0.2206

MNIST Defense CW k=40 randomrestart=1

natural: 98.53%

adversarial: 94.31%

avg nat loss: 0.0409

avg adv loss: 0.1704

All of the above have epsilon=0.3 and step size=0.01

CIFAR10 Attack PGD steps=7

natural: 95.01%

adversarial: 0.00%

avg nat loss: 0.2084

avg adv loss: 42.0510

CIFAR10 Defense PGD steps=7

natural: 87.14%

adversarial: 49.64%

avg nat loss: 0.4592

avg adv loss: 2.9006

CIFAR10 Defense PGD steps=20

natural: 87.14%

adversarial: 45.66%

avg nat loss: 0.4592

avg adv loss: 3.2901

CIFAR10 Defense CW steps=30

natural: 87.14%

adversarial: 46.54%

avg nat loss: 0.4592

avg adv loss: 3.2096

All of the above have epsilon=8 and step size=2.0